AI outputs are becoming harder and harder to read. I don’t just mean their sloppy smell but a lot of the time I just want to scream “speak to me like a normal human”, which, I guess, is ironic.

So I’ve been testing two new instructions. The output’s been 10x better for me, especially when it’s something “I’ve” built. If I say “I’m not technical,” it uses analogies for everything, and honestly I think that’s worse than the slop.

This is what it’s like when describing a Chrome extension I built (which I discussed in Friday’s post*).

You just add this line to your custom instructions (in settings) or your AGENTS.md (it syncs both anyway).

wtf is ‘ASD-STE100 Simplified Technical English’? It’s a writing standard. It makes technical information easier to read, understand, and translate.

I also added:

The two instructions in combination are the sweet spot for me. The ADHD line gives more succinct bullets and headers. It’s a skimmers dream!

*I didn’t realise all the images were transparent so when you enlarged them they looked like shit. I won’t do that again.

Meanwhile my wife thinks there ‘must be someone or a robot’ at the other end of her chats.

Ben’s Bites is brought to you by Cloudera

72% of enterprises say their current data architecture needs a significant overhaul to meet AI goals. See how that shift is taking shape in Cloudera's Great AI Re-Architecture report.

Headlines

OpenAI updated the availability of its models in the good ‘normal’ ChatGPT. Paid users get a special version of GPT-5.6-Sol as the default, and free users get unlimited chats with GPT-5.6-Luna (likely on lower reasoning levels).

Plus, they’ve expanded their cybersecurity program Daybreak and released GPT-5.6-cyber with it. Astra, OpenAI’s next mainstream model, is critical for cybersecurity, so we will have to wait a bit for it to release.

Bloomberg released a scoop on OpenAI’s secret new device - it’s likely a smart speaker without a display in a doughnut-like shape. Easy to carry, over $300 a piece and coming in 2027. (non-paywalled version)

Meta is back to open-sourcing models - Muse Glimmer is their new open-weight model. 30B parameters and tuned for agentic tasks (including computer use). Meta says it’ll soon release the weights of Muse Spark 1.2 - their best-performing model till now. Zuck also wrote a 6000-word piece on his and Meta’s philosophy and values on AI. TLDR: Mark thinks everyone will have

• personal agent that understands them and their goals

• tools for creativity, business and education

• free or affordable access to these tools

Grok has a new image generation model - Imagine 2.0. It is #3 on the image generation leaderboard and #2 on the image editing leaderboard managed by arena.ai.

Your sessions in Claude Code can now message each other.

More Claude news:

• Anthropic will watermark all Claude-generated text (🤮)

• It is planning an IPO in September/early October (I hope my SPV shares count!)

• One of their unreleased Claude models did some more cool maths

Share Ben's Bites

My Feed

• Reliable AI is hard. Nutrient converts PDFs to JSON or Markdown, with coordinates + confidence per value. Grab an API key & try it for free.*

• Raindrop Signals 2.0 - uses custom classifiers to find problems in AI-agent traces with near-frontier accuracy at much lower cost.

• Every company needs a Cassandra - short read, worth your time.

• Agent Plugins - standard for packaging MCP servers and Skills. Supported by Codex, Cursor, Copilot and more.

• Stagehand - the SDK for browser agents.

• Claude Code for normal people with Grace Clark.

• Energy - Another desktop app to work with agents. It creates job-specific assistants instead of using your folders as the home for projects.

• Gloss - prompt any website into looking however you want. Installs as a Chrome extension. (see example)

• seoskill.dev - The only SEO skill your agent needs.

• Replay - Mac app to download videos from YouTube and X to watch later.

• Mintlify Index - search layer for technical knowledge from 200k+ orgs. They claim 2x faster results and 9% higher accuracy vs Context7.

• Lindy Teammate - Let anyone from your team text lindy from Slack like they would another employee.

• The age of “no code” has passed. Or maybe you can build a new version of Zapier. (don’t though, I’d like my shares to be worth something)

• Kitesurf by Cloudflare - a lightweight web browser for AI agents.

• Spotify has a coding app now, Xirp - it combines all your work across Claude Code, Codex, and Gemini CLI in one place.

• Designing with AI? Make a jig, i.e., a tool that makes it easier to make something else.

• AI adoption is a myth.

• Friction as a feature.

• Siri sucks. So I made Codex act as my iPhone’s voice assistant. Here’s how.

• We’re not using enough inference.

Subscribe now

Afters

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I’m trying something new - this email walks through one of my actual agent sessions and I’ll explain what’s happening along the way. The build or task I’m doing isn’t important. But I’m looking at how I could be using agents more effectively.

You might pick up a thing or two, I’m finding it helpful to solidify my own learning.

Please comment what you thought of this, was it helpful? anything unclear? want to see other things?

Subscribe now

What I was trying to do

I was setting up a bookable appointment link in Google Calendar and got annoyed that you can’t just drag time slots on the calendar grid, you have to type each date and time range into a clunky form. I wanted something that would let me drag slots directly on the week view and have the form update automatically.

So I fired up ChatGPT (I use Codex mode but works in ‘Work’ or Claude Cowork etc). I wanted to test Luna on Max reasoning as the price has been cut 80% and people have been saying how great it is to use.

It’s not the best prompt, I’ll admit.

But it gives the agent enough understanding of what I want so it can explore options. Plus a screenshot so it knew what screen I was on about.

This kicks off the ‘agent loop’. The agent thinks about what to do (what can be done with Google Calendar), then acts by using a tool (in this instance, web search) to gather context on how to solve my task.

The websites it read are now in the context window. I didn’t look at them so I have no idea what info it found or if its true. All the text it read is now in its ‘memory’.

Imagine 20 websites went in, there could be wrong or contradictory info that could mislead the agent. This is why you hear so much talk about context. It’s important, and you want it to be full of the best possible information.

Agents often do many loops for a task. They’re ‘go-getters’ by nature. Which is why they need babysitting.

They keep looping, gathering more context until they have what they need to complete the task. For my fantastic prompt, 55 seconds and 2 web searches was enough.

Build it

The agent came back with a mini plan. I skimmed it, as usual (which cost me...).

I overlooked one point which was not how I wanted the extension to work, it should create the times automatically in the form as you drag tiles, not manually click to sync.

What I should’ve done is gone back and forth to ask how things would work, maybe mockup some wireframes I could annotate with feedback.

But I didn’t.

I just said build it...

It cycled through it’s loops and it was built!

Ha, not quite.

My first thought here was:

1. I shouldn’t need to install this myself

2. If its not installed, the agent can’t have tested this live

3. Why did I just say ‘build it’!?

The agent has tools it could’ve used when looping over the task, specifically Computer use and Browser use. It could’ve installed it and tested it live on my actual calendar page.

It didn’t, so I knew there’d be hiccups.

I installed it and tried dragging time frames but the form syncing didn’t work (shock).

What I should’ve done is say something like

That would’ve saved me time and tokens...

I went through my frustration escalation.

• I start by typing the issues I run into.

• 14 minutes later still had issues. I moved to stage 2 - voice ramble and a screenshot.

• 61 (!!) minutes later still had issues. I moved to stage 3 - I record my screen with a voiceover, pointing my cursor at moments with issues. Agents can break videos down frame by frame and transcribe to pinpoint what you’re talking about.

Each turn (back-and-forth) the context gets fuller. After the first set of issues I sent it, it started actually using Chrome to test, thankfully.

But from reading it’s thinking, it didn’t test fully as it didn’t want to override my ‘work’ but that page was for the agent to absolutely use, that’s kind of key to test if this thing worked properly.

So on the second attempt at fixes I rambled a voice note and added a screenshot. This time I added things it should check and tests it should do.

This is that verification layer an agent considers when it’s thinking about the task being complete. Do all the tests (that the agent comes up with) pass with no issues = extension works.

When giving an agent a task, you should think about what criteria would mean this task is ‘done’. For a website it could be that all the content is formatted well with spacing, your design system, and works on mobile. For email triaging it could be that all the emails in your inbox have a label and are moved to the correct folder.

Verification is something I’m still working on as a lot of my tasks are not code.

Development tasks are easier in this respect - agents check ‘did this code run successfully’ (ie. does the code work).

You can also ask your agent to help come up with the verification criteria…

The agent was looping for an hour (!!)... I should’ve known something wasn’t right, but I thought maybe it’s just being really thorough 😅.

It turns out the context window hit the limit 3 times in that turn.

When the context window gets full (~250k tokens), it summarises everything, but keeps some things intact like your messages and the system prompt (the instructions the agent has baked in already).

In this hour, the agent had been testing a lot and finding out things to help it test the extension properly.

But then compaction wiped it from the context so it had to relearn it again, and again.

A bigger model may have helped here as it could more easily recognise what information would be helpful to keep through a compaction. The agent harnesses themselves all compact context differently. Most of the time I don’t worry about it, but this was an example where I should’ve said ‘when testing, note down your approach and learnings’ or something which the next context could’ve picked up.

If it’s important and in your context window, put it in a file! Files don’t get wiped in compaction, only what’s in your session.

I’d also noticed in the agents thinking that it had spent a bunch of time (and tokens) testing an older version of the Chrome extension, it had been saving each new version as a separate folder. Wtf.

This is where I switched models to Sol with High reasoning, Luna Max is fine for everyday tasks, searching, and things like that, but for code and testing it didn’t work well for me here. Skill issue?

I hit level 3 of the frustration escalation so I added a screen recording.

And just to clear up the older versions I followed up with a message mid-task.

Most agents have this in settings. I often send a prompt and then think I better quickly follow up with some extra info or another ask. This sends your prompt whilst the agent is already working, so you can steer it.

13 minutes later it was done. It worked exactly how I wanted it to. Good boy Sol!

So the extension works, it’s useful. I may use it twice a month. But my process wasn’t great - a lot of reminders in there for future sessions.

Bugs are inevitable, for everyone. You just have to think about how you can provide the best context for the agent to figure out its way to fixing them.

What I should’ve done:

1. Read the agent’s mini-plan before saying “build it”

2. Clarify the details upfront instead of discovering it as a bug

3. Tell the agent to install, test, and iterate in Chrome from the start

4. Use the bigger model for iterative debugging loops

5. Be explicit about compaction: tell the agent to preserve key learnings

6. Check the agent is editing and running the same files

None of these are about the agent being bad. The agent was actually really capable — it built a working Chrome extension from a screenshot and a two-word prompt, it controlled Chrome to test its own work, it diagnosed bugs from a screen recording.

The inefficiency was in how I steered it. Every extra hour and every wasted token came from a decision I made (or didn’t make) about what to tell the agent and when.

Here’s the demo video it produced:

Did you like this type of post? Comment and let me know 😊

Behind the scenes

This is how this post came together 😂

Share Ben's Bites

• Find me on X

• Read about me and Ben’s Bites

This time last week I told you I’d downloaded t3 as my ‘all-in-one’ agent app, which lets you use your chat/claude subscriptions from one place.

Well forget that. (strong opinions, weekly held)

There’s a new kid on the block and I’m in love…with the name, bb, but also the app itself.

If you use claude/chatgpt/pi/cursor/factory/[any agent] - you can use it within bb. It looks and works a lot like the chat/codex app (great). I got it set up on my mobile in like 4 seconds which is miles above other apps I tried (t3 was finicky for me here).

But the biggest thing of this app is its extendable nature. There’s a lot of talk on extensible software right now. Pi is one of the best coding harnesses out there and it’s build to be fully customisable (btw is it extendable, extensible, are they the same?!) - the agent can improve its own tool.

• The minimalism of Pi makes it more performant on tasks vs other harnesses.

So much so that companies like Prime Intellect built their agent on top of Pi, as well as many other companies.

• Prime Agent - self-improving harness for coding and long-running tasks for RLMs.

BB is built the same way. It’s an ‘IDE’, except that it isn’t, IDEs are older tools like VSCode that I wouldn’t want to be caught dead in these days. The ‘IDE’ of today is these desktop agent apps that we’re familiar with. BB can extend itself, you want a plugin? Ask it to build one (I got prime-agent to build a factory plugin for it). Want a task tracker? It’ll build one.

I’m never going to spend 100% of my day in an app where I cannot switch models (and agent harnesses), I just cant. I like GPT models for some things, Claude for others. Sue me.

I think the software that we’re going to want (ie builders but not developers) looks like mini ‘widgets’; todo lists, emails, feed readers, etc (like my canvas exploration from last week - which I’m still using btw) and bigger tools like my twitter bookmarks and trends tool (which is private but I may open up one day).

And the tool that you spend most of your ‘work’ time in should have the ability to support adding extra capabilities to it as you learn what you want or need to get your shit done.

Semi-related, there was a few posts on X about why agents aren’t being used outside of developers and early adopters. A quote on the original said:

“You cannot make the masses want an automation flow when they want to hang out w/ their friends and not think about their job.”

I 100% agree. And I still think the pool of ‘builders’ will skyrocket. My definition of that is loose - anyone who wants to use agents to build tools for themselves and/or others, for fun, work, life, whatever! Where do you land in that?

In fact, I’ve believed in this pool of builders for about 10 years now but no-code was too clunky and limited to unlock them, AI is so much easier. It benefits from word-of-mouth. You don’t have to ‘learn’ anything per se, but you start to feel a way of working with it and pick up stuff along the way. Like learning a language, if you’re in it you’ll pick it up faster.

Ben’s Bites is brought to you by Type

Work is multi-player. Now AI is too.

Type puts your team's best AI work into a shared space. Bring your company's knowledge, integrations, and skills together so everyone can use them with any model. Try it now.

Headlines

Bending Spoons is acquiring Airtable, for far less than its last valuation. Though Hyperagent, Airtable’s new AI business, was spun out and is not part of this acquisition. More details on the deal. Another casualty in AI’s wake.

To jump on the news: Obsidian created a free tool to export your Airtable content into Markdown. Another one: Cloudtable - An Airtable clone but it runs on Cloudflare (repo).

Google’s Chief Scientist, Jeff Dean, is leaving the company after 27 years to start a lab of his own with three colleagues. The new company, Discovery Loop, wants to automate machine-learning research. This means a lot of internal moves inside Google. Most notably, Demis Hassabis is stepping away from GDM’s day-to-day leadership to focus on long-term research.

Cloudflare has been releasing a ton of agent-first tools recently: Cloudflare OS is a shared AI workspace with connectors, plus Agent Tracing lets you replay agent sessions and follow them from invocation through model calls. Cloudflare Wallet is another initiative from them to give agents identities, spending limits and wallets for paying for services.

Gemini Notebook (i.e. NotebookLM) now lets you create the different types of artifacts it supports (quizzes, slide decks, etc.) right from the chat window, instead of the side panel.

Meta launched Muse Code, a CLI coding agent powered by Muse Spark 1.2 - a new model with coding-related upgrades, bringing it close to Grok 4.5 level of performance.

Goodfire AI made their interpretability and training tool, Silico, generally available. It’s $1000/mo for individual researchers. These guys tried Silico partly while teaching a 30B model to play a board game, worth reading their experience.

Subscribe now

My Feed

• Capsule: Video Skills and agents for enterprise teams who don't want to sacrifice brand quality.*

• Atlas by WorkOS - an AI teammate in Slack that answers questions, takes actions and learns how your company works.

• Context.dev - API that turns the live web into structured, continuously refreshed data for agents.

• Notion’s results from a survey on 6,000 professionals across 10 markets on how AI is being used at work.

• givemeanode by SF Compute - give Claude, Codex or any agent H100s for training, benchmarks and batch jobs, billed by the minute.

• Power 2026 - primer on power pricing and data centres for understanding AI’s electricity bottleneck.

• Not Diamond Code - model router for coding agents, with claims of 20-65% lower cost on long-horizon tasks.

• Recursive self-improvement is limited by verification, not computation.

• Running Streaks - Opus 5 turned runners’ data into an interactive D3 visualisation using this prompt.

• NYC AI Atlas - open-source 3D map of New York’s leading AI startups.

• Two more cybersec incidents at OpenAI came up in external cyber evaluations.

• Mistral launched Shieldstral - a small 3B open-weight model for content-safety. Easily runs on-device, even on a basic laptop.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I wasn’t prepared for AI hard-hitting truths this morning but I tested this:

In the link, you download the file ‘reflection engine’ which is just a very big prompt basically. Upload the file to your agent and ask:

Please evaluate the attached markdown file and complete all tasks.

I tested it with Fable High and Sol Max, and Sol produced the better report. It was more coherent (agents are speaking more gobbledy goop these days) and nailed connections. Fable’s was harder to read.

I’m going to share some of my report here, because we’re friends - we read and we don’t judge.

So yeah, there was a lot in here and now I’m having a 40+ question ‘grill-me’ session with my agent to address it all and come up with a plan 😅.

I actually found it really telling and helpful. I do have all my therapy transcripts and all sorts of ‘memories’ (read: text in files) that it combed through to produce this.

I recommend you give it a whirl.

Ben’s Bites is brought to you by Nutrient

AI hallucinating? Your document parser is probably dropping fields.

With Nutrient, send any document and get back a JSON schema. Every field returns a bounding box, a confidence score, and a match label, including an honest not_found when values can’t be grounded. 5,000 credits a month, try for free.

Headlines

OpenAI reduced the price of GPT-5.6 Luna by 80% and Terra by 20%. Luna at max thinking effort now scores roughly the same as GPT-5.4 xhigh, the best possible model about four months ago, but at only 8% the cost.

That’s 10-12x more work (or play) you can do if you stick to 5.6 Luna with a little bit of Terra and Sol usage here and there.

In a few tasks I used Luna Max and found its great if it’s not building-related work ie chatting, research, reading/writing files for day to day stuff but I had to tap in Sol High to clean up some mess Luna made on a little chrome extension I made yesterday.

In the meantime, OpenAI also teased their new major model - Astra. They released solutions for 10 long-standing problems in maths and theoretical computer science, all found by Astra. And yes, it’s maths, not math.

Google launched Gemini Robotics 2, one AI system designed to work across everything from robot arms to full humanoids.

It has one model that turns vision and instructions into movement, another for planning complex tasks, and an on-device version. Google says it can control a humanoid’s whole body, handle delicate tasks and coordinate multiple robots.

More models to pay attention to:

• DeepSeek V4 Flash costs $0.14/$0.28 per million input/output tokens—less than Luna’s $0.20/$1.20—with a 1M-token context window.

• Qwen3.8-Max is a 2.4T model that Qwen says comes close to frontier performance for coding and professional work at $2/$6. Its weights, plus those of a smaller 27B model, are coming next week.

• Inkling-Small is open-weight already. It activates 12B of its 276B parameters, and Thinking Machines says it matches or beats the roughly four-times-larger Inkling on many benchmarks.

• MiniMax H3 - a new video model that Krea says matches Seedance’s features for less; open weights are promised soon.

• P-Image-Ideogram - four image-generation quality modes, native 1K/2K output and prices starting at $0.003 per image.

Subscribe now

My Feed

• Ben, who worked with us a couple of years ago & is now at Vercel, built this marketing team template for agents - bring work to a team lead and it delegates the work to a team of marketing agents built with Eve (Vercel’s agent framework).

• YC open-sourced a customisable multi-agent harness it uses across accounting, legal, events and engineering. Here’s the repo. Vercel also built an internal agent called “@v” for finance, comms, docs, marketing, engineering and business analytics.

• ChatGPT desktop app has a new sidebar view - press the bell-like icon, and you can view all conversations that need your attention quickly. Although I don’t understand it at all… I thought it would sort based on recency but it’s not?

  

• GenOffice - free, open-source AI office suite for Mac and PC, built it in a week using $10,000 of tokens.

• The next five years - Patrick Collison’s short survey on how AI will change wages, productivity and the wider economy.

• Conductor Cloud - multiplayer cloud workspaces for Claude Code, Codex, Cursor and OpenCode that keep running after you shut the laptop.

• The human side of AI - taste, fried attention, frozen competence and why more automation can create more human work.

• Cloudflare Computer - open-source virtual filesystem and execution runtime for agents, with container, shell and JavaScript backends on Workers.

• Devtools must be open source.

• The harness is in the app - build the agent loop into the product, instead of adding it from outside.

• The session you cannot take with you - a five-part test for whether an agent session is actually yours: inspect, export, replay, audit and delete.

• How Monologue rebuilt its site with Codex within a week.

• Comp AI’s CRM - open-source, agent-first CRM with durable research agents; MIT licensed.

• FutureSearch - ask for forecasts, including “if I do X, will Y happen?”; its superhuman accuracy claims come from the company. (launch)

• remote-agent-browser - browser-control API for agents with managed cloud sandboxes and no local Chromium install.

Afters

You should go and buy my buddy Jack’s new book - he’s a phenomenal writer and thinker. His angle on anything is always something I look forward to reading. (It’s not only for your twenties…)

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Following on from Tuesday’s messing around building, I built a few more widgets on my canvas. It pulls my X bookmarks, emails, and todos.

Before making this video I didn’t have Loom installed (since it’s crap after being acquired), so just told Codex to build me one.

I drew the image on the left, then just sent that prompt and it worked straight away.

So as software and mini tools are getting easier to create, the tools to create them are not...

I use Codex as my default app because it’s better than all the others and works the best on mobile. But I just downloaded t3 because they basically copied the interface and features, but it lets you choose between different agents; codex, claude, cursor (pi + droid soon). It’s made by a reputable developer that you may have seen mentioned here before, so I trust it’s built well.

At the moment, I ask Codex/ChatGPT to be the orchestrator and to go ask claude about something that is design-related. Which is fine-ish but not great as user experiences go.

What I’m noticing by doing this is how well Codex creates prompts for other agents to follow. I just have a normal chat in my session, then say ‘Use X agent to implement this. Watch its progress and send screenshots every time new work has landed.’ To which it sets up its own monitoring every 5 mins or so and updates me in the same thread, with screenshots.

I’m going to try and put together a ‘bites of the week’ email over the next few days to try and summarise all the stuff going on, what themes people are talking about (loops?!, software factories?!, etc) and explain them.

Let me know if there’s anything specific you need to wrap your head around (I may need to too).

Ben’s Bites is brought to you by Brief

Struggling with fragmented context, slow product decisions, and rework? Brief distills your critical product context into an opinionated graph, then puts a PM agent everywhere you work, (e.g. Slack, Claude Code, email) reducing alignment tax and accelerating cycles. Learn more.

Headlines

OpenAI used Sol to optimise Sol itself, cutting serving costs by 20% and making it 15%+ more efficient at generating tokens. And turns out, it also tops the ARC-AGI-3 benchmark.

Well, there’s a catch: OpenAI says the official ARC-AGI harness hurts Sol’s performance by “forgetting” its reasoning every turn and disabling compaction. Fixing these two things triples Sol’s score from 13.3% to 38.3%, with 6x fewer output tokens.

Re: last week’s fiasco of an OpenAI model hacking Hugging Face - HF published a full replay of roughly 17,600 actions taken by the model. METR and Redwood Research will also independently review what happened.

Though OpenAI is not out of trouble just yet, a Reuters report claims that the same model broke into a customer account at another company (Modal Labs), with rumours suggesting that even more companies were affected.

Anthropic also claimed that Claude Mythos found better attacks on two cryptographic algorithms, though neither affects systems in use today.

Separately (not at all as a reaction to this general trend, right?), ~1300 people working at leading AI companies (OpenAI, Anthropic & others) want the US government to help “pace the frontier” of AI development. Kinda expected when the pace picks up, but this time a lot of the “model makers” themselves are in favour of this pause/slowdown.

btw, The Information reports ChatGPT is nearing one billion weekly users - a milestone OpenAI originally hoped to hit seven months ago. More from OpenAI this week: Codex Security CLI, free frontier access for Academic Researchers, and two new transcription models.

Grok app builder - Grok has a vibe coding interface inside its app now. Create games and apps that can be shared directly to the X timeline. Also see: Drawesome - a zero-dependency drawing toolbar for React, built over a weekend with Grok Build.

Pangram 4 claims it catches 98.83% of humanised AI text with one false positive per ~24,000 docs. An early test found all 38 AI-written words inside a 1,198-word story, though not on every run. Its new image detector claims 99.5% accuracy too.

Subscribe now

Quick links

• Tavus - Build AI that comes to life: video agents that see, hear, and answer in real time and do anything you want. Use TAVUS50 for 50% off.

• 66% of July traffic on docs built with Mintlify was from agents.

• Resend added an MD version of their pricing page to avoid confusing agents.

• 0%, 50% or 200% - ignore AI, halve staff or double the ambition.

• Slackbot can now run code in the background for data analysis, slide creation, and to make live reports or widgets.

• Gemini’s macOS app got a voice mode that lets you ramble, and the app turns it into a clean prompt. Hold Fn to try it.

• The AI future is for everyone - Mark Zuckerberg

• Replit Design - make sites, prototypes and graphics from prompts, URLs, Figma files or screenshots.

• What’s gone wrong with AI & labor.

• Kami - open-source Hermes agents that find customers, prepare outreach and content, then act after your approval.

• Coast - fully local memory for you and your agents, built from what you see on your Mac.

• Pragmatic leverage in the software factory.

• Crew Studio - find useful ideas where agents can help your business, build those agents with the option to take the code home to run anywhere.

• HeyGen Video Podcast - turn a doc, link or idea into a two-host video with scenes, camera cuts and B-roll.

• Copper - local scratchpad for saving answers, links and follow-up prompts across your AI apps.

• FT Chart Doctor - visual vocabulary and examples for choosing a chart that fits the relationship you need to show.

• Mitchell Hashimoto (Ghostty) and Andrew Ng (deeplearning.ai) are both starting new companies: Superlogical and LearnVector.

• MCP’s biggest update removes the need for servers to remember every ongoing connection, making them easier to run and scale.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I’m back from holiday and straight back to messing around building. I’ve been kinda obsessed and enamoured with the demos from tldraw. It’s a free canvas/whiteboard app but their offline app (recently updated) is cool because agents can do all sorts with it; create visualisations to explain stuff, make games, interactive blog posts and a ton of other stuff.

Like turning our logo into a little animating mascot. He’s called ‘bites’ btw.

or little widgets like a timezone checker

or a daily weather timeline

I’m not good using agents when everything is just text in a file. It’s boring, I don’t read it and I sure as shit don’t do enough (anything?) with it. So if it’s visual and all on one canvas, maybe I will?

My idea here is to create a little ‘one-stop’ canvas for some widgets I use regularly as well as showing me my todos/emails to sort and things like that. I’m then planning on making my personal site an interactive canvas too, then some more explainer-like stuff too.

And maybe I’ll get the mascot hooked up to an agent so i can drag him to my todos and ask him to do stuff…

I’ll share more soon 😊

Ben’s Bites is brought to you by Ability AI

This startup runs on 17 AI agents - and open-sourced the platform. Marketing, client health, competitive intel, finance - real ops, not demos, with every agent action git-tracked and a human approving anything external. Agents you own, not rent. Apache 2.0: github.com/Abilityai/trinity

Headlines

Claude Opus 5 is out - Anthropic says it gets close to Fable 5 at half the price. Claire hated using Opus 5, but when asked to rank model outputs in a blind test, she ended up rating it first.

Every’s review lists some behaviours that might explain this: it argues, stops early and fights the skills people built for older models.

And maybe we are overprompting it.

Anthropic says it removed more than 80% of Claude Code’s system prompt for Opus 5 and Fable 5 with no measurable coding-eval loss.

ChatGPT Voice can now run the ChatGPT desktop app. You can use it in ChatGPT Work and Codex both; just start a “Voice” session and speak to it. It can create new sessions for individual tasks and then report the results to you in the main thread. It feels a lot like OpenClaw with voice.

I didn’t have a great time using ChatGPT Voice for “serious” work. The delegation of tasks to new sessions via ChatGPT Voice doesn’t feel as smart as doing it in a chat with 5.6 Sol in Codex. Maybe it’s the lack of control on what models/thinking level is used for each new session.

I imagine Voice to work great for productivity type tasks, checking emails, creating draft replies, pull some info from a dashboard and create a chart etc.

For serious work, I still prefer using a speech to text app (currently using one I built — optionafk.com) and sending that to the agent of my choice with full control over model/thinking effort.

— Keshav

Also, I wouldn’t blame you if you missed it, but Claude also has a voice mode. Till now, it could only use Haiku, but now it can use Sonnet and Opus as well as call multiple tools like Gmail, Calendar and Slack mid-conversation. I don’t remember when I last used Claude’s chat option. It’s been Claude Code for a long time now.

Kimi K3’s weights and technical report are out. It’s already available in Droid at 50% off until 10 August.

Following rumours that the US might ban companies from using Chinese open-weight models, NVIDIA’s CEO Jensen Huang made his first-ever post on X - a long statement backing open-source AI. Over the weekend, everyone signed that statement (OpenAI, Google, Microsoft and 150+ more) - except Anthropic.

Of course, that led to some backlash, and Anthropic did what it does: published an essay to save face, saying it never pushed for an open-weights ban. The reaction to it is people calling Anthropic out on their hypocrisy — they are okay with open models as long as they are cute little 8b toys, but not if they are “big, dangerous models” (to Anthropic’s business).

Subscribe now

Quick links

• FLUX 3 - one model for image, video, audio and action prediction; video is in early access.

• Health in ChatGPT - US users can optionally connect Apple Health and supported medical records, then ask about changes in context.

• BackSearch - search the web as it was on a particular date in the past. Right now, it only works for news in 2026.

• here.now Workspaces - private team home for the sites, docs and dashboards your agents publish.

• Why software factories fail - part 1, part 2, part 3.

• open-git - a code host built for agents, with an autopilot that fixes your PR when CI goes red.

• OpenWorker - open-source Mac agent that finishes everyday work and asks before consequential actions.

• Pilot Protocol - a network of over 250,000 agents that you can use to gain distribution for your app. (I’m an investor)

• Notion as code - define your whole workspace in TypeScript and deploy it through the API. (beta)

• The Stack v3 - open code dataset with about 5T filtered tokens across 770 languages.

• box - full VM sandboxes for your agents at $0.00001 a second, with snapshots, SSH and a desktop.

• Mintlify - turns resolved support tickets into proposed documentation fixes.

• Cognition acquired Interaction - the team behind Poke joins the makers of Devin.

Skills section

• Interface Craft - four new lessons on advanced animation, plus an updated DialKit skill for timelines. (tweet)

• /prototype - asks your agent for several UI versions and saves the chosen one in a shareable URL. (example run)

• Transitions.dev - 33+ polished UI transitions as copyable CSS or React, or an installable agent skill. (tweet)

• sharechat - share a Claude Code or Codex session as an unlisted link; no signup, no login to read it.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Another day in the Vercel vs Cloudflare feud: this time they are fighting over whose AI gateway is faster.

Here’s the result of last week’s poll:

I guess everyone likes Fable more.

Ben’s Bites is brought to you by Metatate

Most agentic data work is quietly propped up. The agent returns something plausible, and every answer gets checked in case it's plausibly wrong. Metatate gives agents the rules they're missing: which revenue definition to use, which policy applies, which records to trust. Try it for free.

Headlines

OpenAI’s models hacked Hugging Face - by accident. OpenAI was testing its models (Sol and an unreleased one—GPT-6??) on a cybersecurity benchmark with safety refusals switched off. The models found an unknown bug in the test environment, and a few more, and eventually broke into Hugging Face’s production servers.

And why? To steal the answers to the test.

Both security teams caught it, the bug has been reported, and both sides have published what they know. Hugging Face says open models were a key part of its defence - its team fought back with GLM-5.2.

Simon’s write-up is always a good read.

Google released some new Gemini models - Gemini 3.6 Flash gives you the same 3.5 Flash performance with a) more efficient token usage and b) a slightly lower cost for output tokens. Gemini 3.5 Flash Lite is a big upgrade over 3.1 Flash Lite, but again, comes at a ~30% price increase. And 3.5 Flash Cyber is a security model for governments and trusted partners only.

There are two things you’d want to use the Gemini Flash models for:

1. Fast speed - if you want a chat-only model with a good context window, it’s a good model to use in your apps.

2. Vision - if you want to use the model for a lot of visual analysis, these are the models you need to pick.

That’s it tbh.

Substack will now tell you what’s AI-written. It’s adding AI detection through Pangram - you can scan posts, replies and comments in the app for an estimate of how much was written by a human.

Pangram has sent the claim of “AI detectors don’t work” for a toss—it works wayyy better than most. But I’m still unsure about how reliable it is. I tested it on some pieces of 100% AI-written content (though that content was a result of a complex pipeline built over months), and I got 100% human scores on most of them.

— Keshav

A relevant experiment: given access to Pangram’s API, Grok 4.5 rewrote an essay 14 times until it passed as human-written, then built a website showing off all 14 attempts. GPT-5.6 Sol and Fable 5 refused to game the detector.

Cursor also launched a router - it picks which model handles each request, claiming 60% lower cost with similar quality of responses. The router lets you select between three options: “cost”, “intelligence” or “balance”.

Routers also have a history of poor performance in real usage. OpenAI’s router, which routed requests between GPT-5’s no-thinking and thinking variants, didn’t fare well. Since then, most companies that build these routers are the ones who sell inference to devs (like OpenRouter), so you don’t really get the feedback loud and clear. With Factory, Ramp and now Cursor making these available directly to users, I hope we’ll get more feedback on whether the routers actually help or if they add too much latency/degrade performance by a lot.

Router or not, we might see more companies adopting this cost/balance/intelligence trio to minimise the headache of choosing the “correct” model for a task.

Claude can now learn a skill by watching you. Record your screen while you do a task, talk through it as you go, and Cowork turns it into a skill Claude can run again - same idea as Codex’s Record & Replay from last month. It’s under “Record a skill” in the desktop app, on Pro, Max and Team plans.

Also: Claude Code got an iOS simulator panel and a security plugin, plus you can now ask Claude about how people actually use AI at work.

Subscribe now

Quick links

• BUZZ - Jack Dorsey’s open-source group chat for teams of people and agents, aimed squarely at Slack and GitHub. (tweet)

• Replit’s mobile app got a full redesign - build and ship from your phone on iOS and Android.

• Slate is a voice journal where the AI never leaves your iPhone - transcription, reflection and storage all happen on-device.

• AFK - macOS app to transcribe multiple-hour recordings without sending any of them to a cloud server.

• OpenAI Presence - voice and chat agents for enterprises that answer questions, use company systems and hand over to people when needed.

• Fable found a 15-30% memory improvement in Next.js’s bundler, nearly autonomously.

• Obliterate, don’t automate - USV on backing AI companies that replace markets entirely instead of making them a bit more efficient.

• YC’s new startup wishlist - AI moving into the physical world: education, healthcare, defence, finance and factories.

• Dana - Applied Intuition’s agentic development environment for physical AI: cars, robots and machines.

• The Claude Code team on how Claude Code gets built - annotated interview.

• Why the team at Factory refunded its first few customers (millions in revenue) before Droid CLI took off.

• Never enough - short post on why AI makes the work rat race feel faster but not more satisfying.

• Devin Outposts lets you run Devin on your own machines - a Mac mini, a GPU box in your lab, or a cluster inside your private network.

• Language Model Builder - build a tiny language model yourself, then chat with the thing you made. (tweet)

• How the Exe team built a distributed DNS server in ~a week and got zero incidents in a month.

Skills section…

• Frontend Textbooks - skill to turn AI’s wall-of-text deep dives into nicely designed HTML books with covers and diagrams. (repo)

• /pick-ui-library - your agent picks a UI library Emil trusts instead of hand-rolling a toast component or installing an abandoned package. (repo)

• and a one-shot prompt to build your own "codex-style" app for Pi, accessible from desktop and mobile.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

It’s my birthday - the grand old age of 36. And I’m on holiday with my family, which, for those of you with young kids, is more full-on than normal life.

There used to be a running joke inside my previous company that when I went away I’d come back with a new grand plan for what we were pivoting to.

Not so much this time.

Last December we had our third child, Poppy, and I was working at Factory. But she has some health issues, and I got pretty burnt out and overwhelmed, so I had to step back.

For the last 7 months I've been plodding along work-wise without much agency. Which isn’t like me at all.

I keep thinking of things I want to exist, but I can't bring myself to put in the work to bring them to fruition. Or more often than not, I'm not happy enough with how they turn out, so I kill them.

Ben’s Bites is the thing I've done longer than anything else. It’s because each email is new. I like new things. I don't like running things. Or doing things because that’s what a business is or should do.

And honestly, with two 3-year-olds and a 7-month-old and being a stay-at-home dad, it’s pretty tough to have the time and energy to shoot for the moon.

I’m going to stop trying to put my square peg in a round hole.

I’m going to put more of myself into this thing. This company is for people who are not traditionally technical but want to use the latest and greatest for work, life and building stuff. That’s who I am.

I’m going to share my opinions on what’s happening in AI more, call out stuff I think is bullshit and try to explain things I’m learning myself.

I’ll be an open book, like I'm being now.

And I’ll test new things too. Please be open with your feedback, but be nice today; it’s my birthday after all.

Ben’s Bites is brought to you by Veridect

Guardrails filter what your AI says. Veridect governs what your agent does — before it acts. Safe moves clear instantly and free; only the riskiest get the full adversarial cross-examination — greenlight, escalate, or block in ~3s, logged tamper-proof. Try to beat it in 60s.

Headlines

• Kimi K3, a new model from the Beijing-based Moonshot AI, surpassed Fable and GPT-5.6-Sol on Arena’s frontend coding leaderboard (and generally came close to these models on other benchmarks as well). It’s a big model (2.8T parameters) and quite token-inefficient. It is half as expensive as GPT-5.6-Sol on a per-token basis but uses twice as many tokens as Sol, balancing out the costs.

  Kimi K3 is an incredible model, but it’s not an incredible value. - Theo

  It works with images, has a 1M-token context window, and its weights will be released openly by 27th July. For now, the model is available on Kimi’s apps and API, but they have paused new subscriptions because they don’t have enough GPUs to serve the model to everyone.

• Fable 5 is now a permanent part of Max and Team subscription plans for Claude with the same condition: you can use 50% of your weekly limits towards Fable. Pro ($20/mo) accounts can use it via usage credits and will get a one-time $100 credit. Also, Fable found an example that disproves the 87-year-old Jacobian conjecture in mathematics. No big deal, right?

• Horizon by Sierra - agents that work towards a business outcome over days or months, moving between calls, texts, email and company systems. Charged for outcomes, not tokens.

• NotebookLM turned three and got a new name: Gemini Notebook. It also got Collections, a new feature to organise notebooks. One notebook can be part of multiple collections.

Subscribe now

Quick links

• Cursor rebuilt SQLite from an 835-page manual using multiple agent swarm setups. Across many configurations that passed all tests, costs varied a lot → Opus 4.8 + Composer 2.5 cost $1,339 vs $10,565 for GPT-5.5 alone.

• Notion AI can now use a browser for you. (in early access)

• Ramp Router - Send each request to a relevant model. It saves 30% in costs for Ramp internally. Comes with an OpenAI-compatible endpoint.

• yoinks - terminal tool for downloading videos from X, YouTube, Instagram and more.

• OpenShip - open-source platform for deploying and running apps on your own infrastructure.

• Lucy 2.5 by Decart AI - live AI model for editing video in real time with more control.

• Linear Loops - recurring workflows that Linear Agent runs for your team.

• Rivet - connect a local agent and design references, then explore several interface directions.

• How much Ridge spends on AI - a non-tech company breaks down whether token costs are actually a problem.

• 12-factor companies - own your context, rent intelligence and teach agents the work before hiring for it.

• Building diagrams to explain databases using Excalidraw charts and Cursor.

• Harness engineering - 12 theses on designing the systems around coding agents.

• Hilos - team chat where people and coding agents work in the same conversation.

Connecting the dots…

• Aaron built himself an SVG editor with Sol instead of letting the model create the SVG in a single go and hoping it turns out exactly as he wants it to be. Now, both he and the model can tweak the SVG (using code or the knobs).

  • Sunil from Cloudflare makes a similar case in his recent post: one document, two hands - you and the agent both should be able to work on the “document” directly inside an app.

  • tldraw released a free, file-based desktop app for its “draw on a canvas” product. Any agent like Codex and Claude Code can use tldraw offline. So tldraw offline becomes one such surface where both you and an agent can work on a single document.

• The Self-Driving Company - Replit’s CEO Amjad Masad wrote this post sharing how agents are effectively in each part of operating Replit.

  • Gumroad is another good example of a self-driving company. It is now mostly run by an AI called Gumclaw. Sahil (Gumroad’s founder) shared that June 2026 was the first month that Gumroad spent equal amounts on AI tokens and human payroll.

  • Good reads from Gumroad/Gumclaw - automating support and TASTE.md

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

It’s been a week with Fable 5 and GPT-5.6 Sol. A quick poll: which one are you using more in your day-to-day work?

Let’s get to what’s new

Ben’s Bites is brought to you by Attio

Introducing Attio: the agentic CRM. With agents and automations that build pipeline, chase signals, and move deals forward, Attio orchestrates your revenue work around the clock. Loved by high-growth startups like Granola, Modal, and Wispr Flow. Start for free today.

Headlines

• Thinking Machines launched Inkling - its first open-weights model. It has a 1M-token context window and works across text, images and audio. It’s not close to the leading open models; most of those are from Chinese labs. Inkling is available on Thinking Machines’ fine-tuning platform Tinker, to use this model as the base for creating custom versions that are better at specific use cases. It’s a fast-growing market…

• GLM-5.2 is currently one of the best open-source models, and many startups are shifting their workloads from frontier models to self-hosted/fine-tuned versions of this model. But it doesn’t have vision, i.e. you can’t attach images in your prompts. This post covers how to add vision capabilities to GLM-5.2 without sacrificing intelligence.

• GPT-Red by OpenAI - an internal model that attacks other OpenAI models to find ways hidden instructions can hijack them. GPT-5.6 Sol was trained with the help of this model to make it safer.

• Google’s OpenClaw competitor is getting more features. Gemini Spark can now edit Docs, read comments in Sheets/Slides and work across sources in parallel; now 50% faster.

• Grok Build, a coding harness like Claude Code from xAI, faced backlash because it uploaded your code to their servers by default. Soon after, Elon announced that Grok Build CLI is now open source.

• DeepMind CEO Demis Hassabis says AGI is probably a few years away and wants a US-led standards body to test the most advanced models before release. Labs would share models up to 30 days early for cybersecurity, biological risk and deception tests - voluntarily at first, with deployment approval possible later.

Subscribe now

My feed

• Yoroll turns ideas, videos, and prompts into playable games you can publish and share. Use the code BBITES: 300 spots.*

• Scan - ask Claude or Codex to make a shareable visual map of your codebase.

• Ramp’s latest spending data tells what SaaS tools did companies pay for in the last month.

• TogetherLink - run open models like GLM-5.2 inside Codex or Claude Code.

• Grandma - a tool that remembers your stack, projects and coding style, then briefs every new agent.

• Thin prompts, thick context - keep prompts and skills light; put the detail in reusable files.

• Reflect rebuilt its notes app from scratch with Fable in two weeks. Reflect Open is free on desktop and MIT-licensed.

• Find animation opportunities - a skill that spots where motion helps your UI and what not to animate.

• Why we stopped using SDKs - direct API calls are easier for agents to debug and keep errors visible.

• Treat AI tokens like headcount - more AI work does not help without clear goals and quality checks.

• Gemma 4 31B got an update making it faster at prompt processing and slightly better at tool use across many benchmarks.

• Long-Horizon Prompting - define success, evidence and failure checks before an agent works for hours.

• Vibe-coded software drifts when teams lose a shared understanding of the system.

• Code is a medium for thought - what we lose now that we don't write the code ourselves, and do we get our flow back?

• A conservationist turned 40TB of public data into a video game.

• A hidden prompt made Claude leak names, employers and security answers from memory.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

OpenAI finally released the GPT-5.6 models to everyone, with many product changes. ChatGPT’s macOS app and the Codex app are now merged, with a new mode called ChatGPT Work. You can update your Codex app and get the new combined app. Codex and Work look similar, with the details fine-tuned for coding-related vs non-coding-related work. There’s a new plugin called “ChatGPT Sites” that lets you build hosted websites with an optional “Login with ChatGPT” feature. It’s nice for sharing but kept annoying me by making everything that I ask on ChatGPT Sites, so I turned it off.

GPT-5.6 series has three models: Luna, Terra and Sol. Each of them comes with 5 thinking levels - light, medium, high, xhigh, max, and a new mode: Ultra mode. Ultra mode basically allows these models to go haywire with subagents.

Now… these models eat up your usage limits much faster when used at higher thinking levels. So don’t rip ultra; first time I’ve ever been nearly out of usage in Codex. I’m defaulting to sol medium for most building/creativity, background agents for harder tasks and day-to-day productivity using luna xhigh.

Some general patterns I’ve noticed:

• Sol is pretty good at UI, but it’s even better once you give it some references. Sol at Max thinking has really good writing, and it’s fun to chat with.

• Terra feels like a replacement for 5.5 with minor improvements in UI and writing skills. It also feels more steerable, so skills would be useful with it.

• Luna has a bit of a mini model smell, like sometimes it doesn’t “get” what you meant in ambiguous prompts. But it doesn’t fail at tasks you clearly define.

Over the weekend, OpenAI reset usage 4-5 times while they were fixing bugs introduced during the app merge, and they have temporarily removed the 5h limit on your usage. Take note of that because you might end up using up your weekly limits in one go.

These models in the Codex app are really good at Computer Use, i.e., self-driving your cursor so it can open apps, click buttons, and use them by looking at the screen. You should definitely try this with Sol medium/high for some small task to see it in action.

Further reading:

• Sol helped post-train Luna.

• Train a model on your iMessages.

• Using GPT-5.6 without burning limits.

Ben’s Bites is brought to you by Voices

Building or training voice models? Stop scraping. Voices has professionally directed, fully consented voice data ready now—part of 100,000+ hours of custom, premium data. Character performances, 43 emotional states, 8 languages. Request a sample today.

Headlines

• Anthropic again extended Fable 5 on paid plans and Claude Code’s 50% higher weekly limits through July 19. This feels like a joke at this point.

• Claude Code got an in-app browser - let Claude open and click through docs, designs or your app. Claude Code Artifacts now support public links and multiplayer editing, including pages made through Claude Tag in Slack.

• Meta released Muse Spark 1.1 - a multimodal model for coding, computer use and agent tasks with a 1M token context window. Meta is also starting to offer its models in the API now, and Muse Spark 1.1 is available at a relatively cheaper price point vs its peers.

• Apple is suing OpenAI and two former employees over alleged trade-secret theft for its upcoming AI hardware. Apple says candidates were told to share confidential information and bring physical parts to interviews; OpenAI says it has no interest in others’ secrets and is reviewing the case.

• Notion Ship OS runs product development from customer feedback to a merged PR in one workspace. Agents triage, route, and summarise work while people make the judgment calls, using Notion’s existing docs and databases.

Subscribe now

My feed

• The future worth building is human - New mission statement from Thinking Machines making the case for many AIs shaped by local knowledge.

• Tend - open-source ChatGPT Work loop for inboxes, hiring or support that learns from your approvals.

• v0 Design Systems 2.0 - import components, Figma and Storybook so generated apps use your real system.

• NameThat - visual dictionary for UI patterns you can see but do not know how to describe to an agent.

• OptionAFK - Private and fast dictation + transcription toolkit that your agents can use.

• shadcn/typeset - one editable CSS file for styling Markdown in blogs, docs and streaming chat.

• Inference AutoTune - claims to distil a frontier model into your own 1-30B specialist in ~2 hours for under $250.

• Cloud Run sandboxes - isolated computers for agents from Google Cloud.

• Hot takes on AI memory - why retrieval and longer context will not solve company-wide memory.

• A software factory that works - scheduled agents create Linear issues; a label sends Claude to fix them. (What is a software factory?)

• Control ideas, not code - agents are good at local code; humans should own the design, testing and direction.

• AI’s biggest winners may have the lowest margins - agents can cut coordination costs in physical businesses.

• The reverse information paradox - Satya Nadella thinks firms should own the corrections and feedback that make AI useful.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Anthropic extended access to Fable 5 via Claude subscriptions until July 12. Same deal as before: up to 50% of your weekly limits, and after July 12, use prepaid usage credits. But this dangling-a-carrot move doesn’t make me trust their deadlines anymore.

Otoh, OpenAI announced that GPT-5.6 models (Sol, Terra and Luna) will be live for all users later today. General feedback on X (from early testers) is that Sol is not smarter than Fable but is so much more reliable that you would not want to use anything else.

Well, let’s see.

Ben’s Bites is brought to you by Flodesk

Most AI tools generate emails for you. Flodesk Studio designs with you — describe what you want, get a beautiful on-brand email in seconds, then refine in chat or perfect by hand. Real designer foundations, AI-accelerated, finished by you. Try it.

Headlines

• We have a good Grok model - Grok 4.5. SpaceXAI and Cursor trained it together, and it’s an Opus-class model. It should perform roughly in between Opus 4.7 and 4.8. But it’s much cheaper on a per-token basis and token-efficient—making it 6x cheaper than Opus models and 3x cheaper than GPT-5.5. It’s available in Cursor with increased limits and via API too.

• ChatGPT Voice is finally getting some new models - GPT-Live-1 and Live-1-mini. A new architecture lets these models a) talk while you’re still speaking without cutting your message off and b) use GPT-5.5 in the background. Results in much smoother conversations and a big jump in intelligence from the old advanced voice mode.

• A lighter version of Claude Cowork will work on web and mobile - sessions will run remotely and save to your account, so you can kick off a task on the laptop, close it, and check in from your phone. Cowork on the desktop app remains the full deal. Beta rolls out over the coming weeks.

• Muse Image and Video - New multimedia generation models from Meta. The images have more of an AI look, but the model seems good at image editing and text rendering, and can search the web to get facts right, etc. They also have a video model in preview that will be released soon.

Subscribe now

My feed

• Render Workflows - Define long-running tasks with a lightweight SDK that handles queuing, orchestration, & retries. Try it - use code RENDER-BENSBITES for $50 in credits.*

• Cognition released SWE-1.7 - its strongest model yet, near the top frontier models at lower cost, and available at 1000 tok/s.

• Cloudflare Drop - drag a folder into the browser, get a live site, no account needed.

• Toya - Executive assistant that lives in iMessage, calls your phone, knows you and your work. (i’m an investor; check them out on Product Hunt and Twitter)

• Ora Directory - search the web for tools and resources agents can use.

• 64 Claudes rewrote Bun in Rust in just 11 days.

• Spatial context in Clicky - draw on your screen so the model knows exactly where to look.

• LoginWithChatGPT - unofficial SDK to add “Sign in with ChatGPT” in your app.

• How to become an applied AI engineer.

• Model vs effort in Claude Code: “big model low thinking” or “small model high thinking”, what should you choose and when?

• How to turn raw Markdown notes into an LLM knowledge base.

• A collection of small workers hosted on Notion to sync Twitter bookmarks, Readwise, read webpages in markdown and more. (repo)

• Perplexity Teammate - reportedly an internal coding agent like Claude Code/Codex.

• Skills collections:

  • mattpocock/skills v1.1 - /wayfinder, /research, /implement, /to-spec and /to-tickets.

  • Meng To’s Agent Skills - 75 design, motion, WebGL and landing-page skills for agents.

  ---

  Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I’ve been using Fable quite a bit since it came back (but not for long…) and I’ve barely used it to do any building. I like Anthropic’s models mostly for their thinking, they’re resourceful, creative and can connect things I often can’t put my finger on.

So Fable’s been a bit of a therapist/thinking partner this past week. But it feels slightly different this time. It could be me, but it has shown me a bunch of Opus-like traits in how it’s talking or thinking. And it’s not because my prompts are being routed to Opus (I checked).

I’m excited to see GPT 5.6 come out, I think this week, to see if they’ve finally cracked the creativity/thinking vibe I’m after. Codex is a materially better product than Claude Code (or Cowork). And it’s super fast.

As usual, I’m using a square peg for a round hole - a coding agent with a coding brain for creative thinking and brainstorming. But I do change the system prompt of Claude Code to a simpler one that is much faster and much less steered.

I’m wanting a harness that can still build but is less of a developer at heart, and more something to really spar with me on creative thinking. I sometimes flit to plain Claude for different perspectives, but that doesn’t seem to cut it either.

I’ve tried skills, prompts, and harnesses but nothing seems to get me what I want. Which is something I keep thinking through when I want to make something…what do I actually want?! Today that question comes up for me a lot, and I keep convincing myself I’ll only know when I see it, so I need to stay ‘in the work’ before it’ll present itself.

But equally, this could be AI having an impact on me - if anything can be generated at any moment, just cycle through until you see what you want. Just keep spinning that wheel.

Seems healthy? 😅

Ben’s Bites is brought to you by Adobe Firefly

Adobe Firefly expands agentic capabilities with: brand kit creation (describe your style, brand name, and color palette and let it go to work), short product video creation, Quick Cut (auto-assemble video footage into a polished cut), storyboards (visualize and sequence ideas before production).

Headlines

• Today is the last day of Fable being available in your Claude subscriptions (atleast for now). You will have to buy usage credits for using Fable starting tomorrow. So everyone has a guide to get the max out of this one day:

  • Use Fable to find your unknowns - Thariq (from Anthropic)

  • Let Fable use its judgement - Simon Willison

  I’ve mostly used a single long chat with Fable this entire week. The way it uses memory, compaction, and writing to files is amazing. I’m making it use codex and droid as subagents A LOT. So much that I’m running out on my Codex usage faster than my Claude Code usage. — Keshav

  Theo’s video on how he has used Fable is a good watch to reimagine how you should use the newer generation of AI models, because remember, GPT-5.6-Sol will likely become accessible this week. You can expect similar performance/behaviour from that one.

• The Information reported that OpenAI has found a way to cut its inference cost in half and already tested it on logged-out ChatGPT users. Does this mean it can reduce API costs of their latest models too? Not sure, but I hope it is. The last major price cut we had was about a year ago, when OpenAI reduced the cost of o3 by 80%.

• GPT-Realtime-2.1-mini - new model in the OpenAI API that can take in video and output audio while doing reasoning and tool calls.

• MIRA is a playable world model trained on 10k hours of Rocket League-style gameplay. It feels like you’re actually playing the game on low-graphics with janky controls.

• New interpretability research from Anthropic - Claude uses a global workspace to activate a bunch of concepts that don’t appear in the output or the chain of thought but affect its work. The researchers are comparing it to unconscious thoughts.

Subscribe now

My feed

• OpenScience is an open-source Claude Science alternative. Works with multiple models and 250+ skills.

• You can run /cd in Claude Code to move a running session to another folder without restarting.

• The Making of Claude Code - short history from the team and early users.

• integrations.sh - catalogue of product MCPs, APIs, CLIs, and auth steps for agents.

• Taste Skill + Impeccable - design skills for agents that keep UIs from looking generic.

• Boring Computers - spin up a real computer for an agent, then destroy it.

• Chromeless - MacOS browser without chrome for clean screenshots and recordings.

• Vercel MCP and CLI can discover and fetch agent traces from sessions with Eve, their own agent framework.

• Types of loops and how to use them - turn-based prompts, /goal, /loop, /schedule and proactive routines.

• exe.dev now lets you connect your ChatGPT subscription to their terminal agent.

• Dayflow logs your screen activity and creates an automatic work journal based on what you’ve spent your time on in a day.

• PersonalOS - 300k-token life context pack from iMessage, notes, docs and app exports. (prompt to do it for yourself)

• Headroom - Mac menu bar app for Claude/Codex usage and spend during Fable week.

  ---

  Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

When I was in Greece, I needed to organise a taxi from the airport home. I pinged codex and 1m37s later it was done, I got a text to say the car was booked.

It’s not the most impressive task by all means but there’s a system as to why it works (and this system is how a lot of tasks work).

So first, all my day-to-day agent conversations start in my ‘bites’ folder, which has links to files that store my memories, what I’m working on, ben’s bites, about me etc.

Codex loads my AGENTS.md and my task then gets to work.

It’s got the default Google Calendar and Gmail tools connected, so starts with looking at if my flight is in my calendar (which it was, thanks to another Codex thread I did pre-trip). Finds the trip and then looks in my emails for any prior task/transfer emails, finds one to get my home address and the company I used on the way out (these could also be noted in my folders’ files, in a ‘TRAVEL.md’ file, linked from my AGENTS.md).

So now it’s got all the details it needs, this is the ‘agenty’ bit, it’s gathered enough context to complete my task. With that context it determines the next thing to do: make the booking. It knows the tools it has available to it, like computer use, so it went on to the taxi’s website (I had my computer stay on all trip so I could connect), filled in the booking form and made the payment.

Working with agents is a lot about making sure the right context and tools are available to do the task. If both are covered, agents can get stuff done without you interfering.

We have to think about tasks in this way - what would an agent (or a person) need if they had no knowledge of you or your task? It’s your job to make sure it’s available.

Let’s get to it.

p.s. I wanted to shout out my friend’s first business he launched solo - Fixxa - voice quotes and invoices for UK tradespeople, with WhatsApp delivery and Stripe links.

Ben’s Bites is brought to you by Attio

Introducing Attio: the agentic CRM. With agents and automations that build pipeline, chase signals, and move deals forward, Attio orchestrates your revenue work around the clock. Loved by high-growth startups like Granola, Modal, and Wispr Flow. Start for free today.

Headlines

• Fable 5 is back for all paid users. Anthropic’s blog post mentioned stronger guardrails in this re-release, but I haven’t hit any of them yet. Fable 5 is only included in the subscription plans till July 7 and you can only use 50% of your limits on Fable. This benchmark claims Fable can do 16% of remote work projects, double the amount of Opus 4.8.

  • Theo has quick tips on getting the max out of Fable in the next 5-6 days.

• Before Fable came back, we also got Claude Sonnet 5 - it’s benchmarked close to Opus 4.8 on most agent tasks, and cheaper on a per-token basis, but in practice it costs roughly the same as Opus per task. It is now the default for Free/Pro, available in Claude Code and API, and has launch pricing of $2/$10 per million tokens until Aug 31. Mine and others’ experiences with it is it’s expensive and slow, I can’t see it being something I use over other models.

• Two new Gemini media models: Nano Banana 2 Lite and Gemini Omni Flash. Both models are available in the Gemini app and API. In the API, Nano Banana 2 Lite gives you fast (under 4 seconds) and cheap images (~30 images at 1K resolution for a dollar). Omni Flash lets you generate and edit videos at $0.10/sec.

• Bridgewater and Thinking Machines trained a specialist model. It hit 84.7% accuracy on financial triage at 13.8x lower cost than the best frontier model tested. Factory’s Droid Shield 2.0 also fine-tuned two detectors to catch exposed secrets in sessions and reduce false alarms.

Subscribe now

My feed

• Modelence Mobile Builder - build native mobile apps by chat, using the same Modelence auth and backend. (portfolio company)

• Browserbase Agents - one prompt or API call for browser automation in a hosted harness.

• Safari MCP - let agents use Safari to open tabs, debug pages and catch Safari-only bugs.

• Option AFK - Wispr-flow alternative with local transcription (nothing leaves your device), ability to transcribe multi-hour files and a CLI for your agents.

• Claude Science - Mac/Linux workbench for scientists with code-backed figures, 60+ scientific skills/connectors and more. In beta for Pro, Max, Team and Enterprise.

• Ramp PorTAL - move fine-tuned tasks between models at about half the usual cost.

• xAI Voice Agent Builder - no-code Grok Voice agents at $0.05/min.

• plain-writing-skill - rules for agents to write plainly, with an HTML diff of what changed.

• Bond - AI chief of staff for founders and execs that tracks context, blockers and follow-ups.

• Interfere - watches production, investigates issues and fixes problems before users notice.

• AI-native PM leverage - how PMs move from text help to prototypes, PRs and evals. (course)

• /wizard - skill to build interactive CLIs for annoying setup tasks.

• Understanding agent-written code - a case for still knowing what the code does when agents write it.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Etched.

Training got all the attention for the last few years. But in 2026, the real fight is serving intelligence: lower latency, lower cost, less power, more tokens. The market is screaming for better inference systems, and we’ve never seen a market like inference. So much so, people can’t sleep thinking about it (see the tweet in afters 😂).

AI inference could be the biggest market in history and the bottleneck is inference hardware. Etched built the perfect product. Pulling it off will change how models run.

Etched are building frontier inference clusters with extreme vertical integration: chips, racks, software, manufacturing, and production co-designed end to end. First product in under three years, while most hardware peers took 7+.

Their traction is already insane: $800M raised, $1B+ in backlog orders, production underway with TSMC, and first chips working on the very first try (A0) on TSMC 4nm. A0 means the first physical version of a chip/system design that comes back from the fab. So yeah, A0 success is hard. Doing it this fast is wild.

I backed Gavin in 2023, as a 21-year old Harvard dropout, because he had first-principles technical conviction and the ability to recruit the most insane team who had actually built the last generation of compute infrastructure. Now Etched is 400+ people from NVIDIA, Google TPU, Broadcom, SK Hynix, TSMC and basically every serious AI chip program.

They’re backed by VentureTech Alliance (which has a strong partnership w/ TSMC), Peter Thiel, Jane Street, Two Sigma, Jump, HRT, Stripes, Ribbit and more.

As we’re seeing, who wins in AI won’t just be decided by who has the best models. But by who can serve them.

Ben’s Bites is brought to you by Render

Chainable compute. Right on queue.

Define tasks with Render’s lightweight SDK and chain them into long-running, distributed workflows. Launch your agents and batch jobs on demand. Render Workflows handles queuing, orchestration, and retries.

Try it: Use code RENDER-BENSBITES for $50 in credits.

Headlines

• OpenAI has GPT-5.6, but its launch is also blocked by the US govt. Only “select partners” are getting access to GPT-5.6 Sol, Terra and Luna, three new models in the family. Sol is the biggest/smartest one of the bunch. It bypasses Mythos in certain benchmarks but is conveniently a little shy of Mythos-level in exploiting cybersecurity bugs. Sam Altman claims that GPT-5.6 will be available to regular users soon—though it might be US only to begin with, even if he’s working hard for worldwide.

• OpenAI also released an economics paper based on the adoption of Codex inside & outside OpenAI. Non-technical adoption is quickly matching engineering departments. The lead of the Codex app was on Lenny’s podcast, where he talked about designing the app and the features that matter.

• Cursor for iOS lets you launch always-on cloud agents from your phone & remote-control agents running on your computer. Composer 2.5 is 75% off in the app through July 5.

• X launched a hosted MCP so Grok, Cursor and other MCP-compatible tools can connect to the X API and X developer docs without building the server yourself.

• How do you build AI that actually understands you? Working Smarter, a podcast from Dropbox about AI and modern work, is back for season three. From context engineering to multimodal search, hear how engineers are building AI that works wherever you do. Listen to the first episode now.*

Subscribe now

My feed

• Replit has a desktop app now for both Mac and Windows.

• US national design studio created Rampart - a 14.7MB in-browser ML model to redact PII before it’s sent to a server.

• In-app custom agents work for regular users; power users want skills they can plug into their own agents to use your product.

• Human in the /loop - how do you reliably get in and out of the way of agents?

• Zaro - build live apps, agents and workflows from Slack, email, docs and calendars.

• Unpeel - native Mac terminal for agents with persistent sessions and git worktrees. (demo)

• Tau - educational agent harness for building TUIs, extensions and harnesses.

• Inference.net lets you test GLM 5.2 on mirrored prod traffic before switching models for your live users.

• Odessia Travel - opinionated AI trip planner that searches and books flights, stays and activities.

• smolmachines - spin up hardware-isolated Linux microVMs.

• Animation vocabulary - skill for asking AI for motion with words like morph and rubber-band.

• Every time you think you need a dashboard, ask your agent to make a throwaway HTML.

• Teaching agents product design at Vercel.

• MCPs, APIs and CLIs - These are not three separate concepts.

• New set of shadcn components for building chat interfaces - streaming chat, scrolling, messages and attachments.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Keshav here.

About a week ago, I watched Codex automatically generate images with the Image Gen skill while making an app. It used them as real assets in the UI.

Since then, I’ve started explicitly asking it to create images whenever I’m building web UIs. The results are noticeably better: pages look less bland, and because Claude Code can’t generate images, the output feels different from the usual AI-generated UI.

Give it a try.

Ben’s Bites is brought to you by Rippling

Spinning up agents is easy. Letting them access your company data safely isn’t. Join Box CTO Ben Kus on June 30 for a live webinar on the guardrails, eval process, and onboarding guide your agents need — including the diagnostic tool Rippling’s own AI team uses — save your spot.

Headlines

• Claude Tag lets you mention a shared instance of Claude Code like agent across your team in Slack. Tag it into work, let it keep context from Slack, and delegate tasks while you do something else.

• Gemini 3.5 Flash has computer use now. It can control the browser, mobile and desktop environments, and Google has a GitHub repo to try it locally or through Browserbase.

• New from Figma Config - turn design layers into code, new tool (Figma Motion) for motion design work, generate editable shaders, vibe coding for plugins, third-party connections for Figma Agent and more.

• Notion’s new developer platform is adding the ability to run code-based workflows and the ability to integrate external agents like Claude Code, Cursor, Codex, etc., so that they can work from shared docs and task boards.

• OpenAI built its first AI chip, Jalapeño, with Broadcom. It is made for the LLM work behind ChatGPT, Codex, the API and future agent products.

• Build & ship at the Runpod Flash Hack Day! Join Runpod on June 30 at the SF Builder’s Collective for an in-person hackathon. Remote-friendly. Learn how to use Runpod Flash to turn Python functions into auto-scaling, serverless GPU endpoints without Docker. Demos, prizes & mentorship. Register here.*

Subscribe now

My feed

• GPT-5.5 Instant got an update that makes it more fun to talk to, better at intent, constraints and recommendations.

• Exa Connect - web agents to query ZoomInfo, Crunchbase, Similarweb and more.

• Perplexity Computer for Counsel - legal research, docs and matter tools in Computer.

• AssemblyAI Universal-3.5 Pro Realtime - speech-to-text uses the agent’s side of the call as context.

• Modal Auto Endpoints - run open models in production with one command.

• Executor - open-source gateway for connecting agents to services.

• Aside - AI browser with vertical tabs, local encrypted data and Claude/ChatGPT support.

• Genspark Design - generate UI prototypes, videos, HTML animations and code.

• Hubble - Markdown notepad for you and agents, with live HTML previews.

• John Jumper, AlphaFold lead, is leaving DeepMind for Anthropic.

• Engram - a new lab that hopes to train a personal model that learns from your work and updates roughly every day.

• Emil’s design skills repo - design engineering skills with 100k+ installs.

• Harvey Labs - legal foundation models, open evals and firm-owned intelligence.

• Codex workflow tip - have your agent write workflow papercuts to /tmp while it runs.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I’ve been looking at a lot of design resources for the manual recently. Interface Craft by Josh is one of the best places I’ve found for learning how to actually feel and notice taste.

Let’s get into what happened over the weekend.

Ben’s Bites is brought to you by nexos.ai, the unified AI platform

Tired of using different AI models for different tasks? We were too, so we made nexos.ai. It gives you access to 100+ models in one platform. Oh, and no more Googling “which AI should I use for this?” Just let nexos.ai automatically pick the best one. Give it a try (50% off with code benbites).

Headlines

• Codex Record & Replay - show Codex a recurring workflow once, like filing an expense report or submitting time off, and it turns the demo into an inspectable, editable skill you can reuse.

• Claude Code has Artifacts now - HTML pages with some functionality that you can share with others (like a PR walkthrough or living project dashboard). Available in beta for Team and Enterprise plans.

• OpenAI is expanding Daybreak, its cybersecurity program, with a new version of GPT-5.5-Cyber (trusted partners only) that can reproduce even more bugs than Mythos and Patch the Planet, a push to help maintainers fix vulnerable open-source software faster.

• Sakana AI released Fugu - An API that picks and coordinates several models for hard tasks. Fugu Ultra claims 73.7 on SWE-bench Pro and 82.1 on TerminalBench 2.1, roughly Fable-class, but you will feel gaps in real usage.

• Build & ship at the Runpod Flash Hack Day! Join Runpod on June 30 at the SF Builder’s Collective for an in-person hackathon. Remote-friendly. Learn how to use Runpod Flash to turn Python functions into auto-scaling, serverless GPU endpoints without Docker. Demos, prizes & mentorship. Register here.*

Subscribe now

My feed

• The coming loop - this is the clearest explanation/writing I’ve read on the loop discourse.

• Cursor /automate - describe an automation; Cursor sets triggers, tools and instructions.

• Claude Code steering guide - where to put instructions, skills, hooks and subagents.

• Gemini’s Interactions API is now generally available. It combines their API for models and agents.

• GPT-5.5 Instant got better at health questions - now on par with OpenAI’s best Thinking models, per OpenAI.

• birdclaw.sh - local Twitter archive with search, inbox and ranked triage. (see more)

• Lettera - native file-based Markdown editor for Mac.

• Perplexity Computer has a new memory system - Brain.

• You’re spending too much on AI. You’re also using it too little.

• Stripe Directory - agents can search and pay Stripe businesses from the CLI.

• Clips - open-source Loom alternative agents can inspect from a URL.

• Can Claude get a robot to play ball? Not fully, but Opus 4.7 was about 20x faster than last year’s Opus 4.1 team.

• Ports - Mac menu app for local dev servers, ports and kill/open buttons.

• Slackbot can now connect to 20+ apps like Linear, Replit, Canva, DocuSign, Zoom and more.

• ElevenLabs Ads Engine - localise ads into 50+ languages and publish to ad accounts.

• 15 copy-paste loops people are actually running.

• Joining a Software Factory - how Factory builds repos for agent PRs.

• Post-Agent Companies - agents make labour cheap; context and trust stay scarce.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Heading to a Q&A with Sam Altman later today in London so a bit rushed for time to part any ‘wisdom’.

I’m currently still mid-building the reference manual, but it’s coming along now! I’ll release a couple preview lessons hopefully by next week.

I’m using Codex as my ‘workhorse’ agent. Mostly because it’s easiest access on mobile while plugged into all my local files. I’m using it for a bunch of automations, check my inbox and triage (rip my little email build), ingesting a bunch of my twitter bookmarks to summarise all the stuff I’m saving and organising it into topics in my memory system - so i can ask ‘what do we know about agent memory’ and it’ll pull from all the posts i’ve saved.

I spent a lot of time talking to agents less, unless I was actually working. But I’m now leaning back into just talk to agents about everything.

Ben’s Bites is brought to you by Attio, the AI CRM

Attio is the CRM for the new way of GTM. Get agents working on every account, surfacing opportunities, and handle the work that used to take your team days. Open your inbox, the follow-ups are drafted. Walk into a meeting, you're already briefed. Got a question, just Ask Attio. Start for free today.

Headlines

• SpaceX is acquiring Cursor for $60B in an all-stock deal. Cursor also launched a few new things at Compile, their first conference (it had chalkboards on stage).

  • A GitHub alternative in waitlist - Origin for code storage/git hosting.

  • Smoother transitioning between local and cloud agents.

  • And they teased a new Cursor model that’ll do much more than coding.

• Midjourney (the image generation company) is building an ultrasonic body scanner under a new division called Midjourney Medical, and a Spa in San Francisco.

• Claude Design now follows your design system, lets you edit the canvas directly & syncs with Claude Code plus more tools like Replit. On the design tool theme: Framer now connects to external agents like Claude Code and Codex, and v0 has a new design mode.

• OpenAI’s financials got leaked - Ed Zitron says audited 2025 financials show $13.07B revenue and $34B costs, with Jack Raines adding useful context on the margin debate. Also, Noam Shazeer, Google’s Gemini co-lead and Transformer co-author, is joining OpenAI.

• As part of the Spring '26 Edition, Shopify now lets any developer build end-to-end agentic commerce experiences. Use Catalog API to query billions of merchant products, and the Universal Commerce Protocol to power the full commerce journey from product discovery to checkout. See what else is new.*

Subscribe now

My feed

• Like insurance for cloud spend: Archera insures AWS/Azure/GCP commitments; reservation savings without the downside. Start with $0 platform fees.*

• Copilot Cowork, the latest tool from Microsoft, faces the same issue as Claude Cowork. It forces an artificial choice between chat and work, while Codex lets a chat naturally become the work thread.

• Grok Imagine Video 1.5 - sharper image-to-video with better physics and speed.

• Computer use, Chrome extension, memory and Chronicle in Codex are now available in Europe as well.

• Tacit Labs - applied research lab for AI and biology.

• AutoWiki by Factory AI - Generate a structured wiki for your code that updates on every push.

• Block built a tool called Builderbot that coordinates agents across our entire codebase. Here’s how.

• Side-by-side comparison of website designs built by Fable 5, Opus 4.8 and Kimi K2.7 (read more).

• Exa Agent - web research API using cheaper model orchestration.

• Ploy - AI marketing platform for websites, SEO, CRM and campaigns. Founded by Webflow’s founding CTO.

• HumanLayer - agentic IDE and collaboration layer for software factories.

• API for Cursor - local MacOS app that lets you use Cursor’s models with any harness.

• Vercel launched Eve - an agent framework that they hope is Next.js for agents. Also see Flue (which hopes to be Astro instead).

• Three ways Codex can use a computer.

• killedbyopenai.com - graveyard for things OpenAI killed

• /visual-plan - visual plans for Codex/Claude Code.

• Polar is moving away from Tailwind to an LLM-safe design system.

• Local models are good now.

Build logs

- by Keshav

Before Fable was removed, I used it to make a tiny CLI utility for moving files between my Samsung S24 Ultra and my M3 MacBook Air.

I just wanted to plug the phone in with a cable and move photos, screenshots, PDFs, etc. without downloading another random app or uploading everything over WiFi.

It suggested two options, guided me through the setup (which only took 30 secs), and then wired everything by itself.

The only hiccup was the terminal command it picked for this utility: droid, which was already being used by Droid, the coding agent from Factory AI.

I renamed it to phone, and now I can search photos, videos or any file on my phone from my Mac terminal, and transfer files both ways: my phone to my Mac, and my Mac to my phone.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Fable went from “new best model” to “not available” in about 3 days.

Anthropic launched Claude Fable 5 on June 9. Fable was the Mythos-class model for general use but with guardrails. Mythos is Anthropic’s new model only available to select companies because they say it’s a cybersecurity risk.

Then on June 12, Anthropic posted that the US government got spooked by a jailbreak found by Fable (which Anthropic claims that other models like GPT-5.5 can also find). The US govt. suspended access to Fable 5 and Mythos 5 for all foreign nationals, inside or outside the US, including foreign Anthropic employees.

So Anthropic’s answer was basically: we can’t do that cleanly, so we’re turning it off for everyone.

Worth reading:

• The government’s side, via David Sacks.

• Armin’s “Americans only” post - if a model is apparently too dangerous, why is the boundary nationality?

• Anthropic is losing the mandate of heaven

But I really like this part from shadcn’s post -

treat intelligence as borrowed. Use the best model while you have it to make durable plans, specs and implementation notes, then execute later with something cheaper or something you control. - shadcn

Worth reading Simon’s review describing Fable as relentlessly proactive.

Anyway, bye bye Fable, I guess.

Ben’s Bites is brought to you by Working Smarter from Dropbox

Modern work can be frustrating and chaotic—if you don’t have the right tools. From context engineering to multimodal search, Dropbox engineers share how they’re building AI that actually understands you, so you can focus on the work that matters most. Listen now.

Headlines

• Software Factory - Factory 2.0 launched and there’s been a lot of chatter recently about building ‘software factories’ - ie build the machine to build the machine, or more simply, create the systems, infrastructure and workflows to produce software with agents. Factory’s co-founder Matan was on the 20VC which is worth a watch.

• Ramp SWE-Bench: coding benchmark created from real engineering problems inside Ramp. Fable 5 > GPT-5.5 ~ Opus 4.7 > Opus 4.8, but each increment in performance comes with a 1.5x the cost.

• Vercel has a new tool, drop.new - It lets you just upload a folder, a zip or a file, and it deploys that as a website or an app. Clearly a way to sidestep GitHub.

• OpenAI now lets you save your Codex rate limits reset. You can then use it anytime in the next 30 days.

• The Information just reported that DeepSeek has raised $7.4B, valuing it at more than $50B. It’s their first funding round, and the CEO himself is the biggest check (~40% of the total funding).

• AI chips keep getting faster, but the infrastructure developers rely on hasn’t kept up. In this piece, Runpod (aka the AI Developer Cloud) explores why the real bottleneck may be the stack around the hardware & what a developer-first approach to AI infrastructure looks like: https://fandf.co/4vQDUT4*

Subscribe now

My feed

• Extend CLI - Give your agent the ability to parse and extract an entire corpus of documents.

• Efecto - design tool with code under the hood so that agents can use the same canvas as you.

• Figma’s Chrome extension now lets you copy/paste websites into editable Figma layers.

• A guide to mastering Codex on Mobile for your engineering work.

• It’s a mistake to call oneself a “non-technical founder”

• Making a human → 10,000 agents “org” work with recursive agent systems.

• Fusion API by OpenRouter- Get multiple models to work on your tasks with a judge model to decide the final answers.

• How I work with coding agents like a caveman.

• React security doctor - Fix security exploits in your React app.

• Sonic-3.5 and Ink-2 - New text to speech and speech to text models from Cartesia.

• A frontier without an ecosystem is not stable. - Satya Nadella

• Castform - bring your data and train open-source models on them without worrying about how to make GPUs work.

• How Fable was probably built.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @sdmat123

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Like a lot of others, I’m trying Fable. But I don’t think I’ve given it really hard tasks for me to feel the big step change people are claiming. Although as I’m working on visual, interactive essays for my reference manual, I’ve been testing one-shot generations on all sorts of topics to play with my kids; how to play the piano, how tetris works, how a volcano works. They’re loving bashing away at my keyboard to see the components come alive.

It’s less chatty than Opus. I use agents as a back-and-forth, so I need some chattiness, GPT models are very to the point, Claude models are usually more verbose. So I think Fable’s a bit of a sweet spot for me, because as of now I still prefer the ‘vibe’ of how Claude models think with me.

It’s slow. Speed is a huge thing for me because I can’t actively run many agents at once, my ADHD context switching is already bad enough. But I need the speed. So then I look to change from High or XHigh reasoning to lower levels but it feels wrong to pick less intelligence. Even though I know it’s probably wasting tokens.

Composer 2.5 Fast from Cursor is the fastest model I’ve tried and I really enjoyed using it in Pi, and it’s shattered the glass for me, GPT 5.5 has too to be honest. They’re quick so I can chew through a bunch of work quickly.

Ever a game of toss-up between price, speed and vibe.

I’m sure GPT 5.6 or 6 is just around the corner, and I imagine OpenAI are really trying to make sure the vibe is on par with Claude models if they want to keep the Codex hype going.

Ben’s Bites is brought to you by Plaid

Over half of Americans used AI to manage their finances in the past 12 months. And their expectations for financial products have never been higher. Plaid's latest report, The State of Intelligent Finance, breaks down what consumers expect from AI and what that means for your team. Get the report.

Headlines

• Fable 5, the new model from Claude, is the “safer” version of Mythos (Anthropic’s unreleased model which they say is a big cybersecurity risk and is only available to certain companies). Fable makes a big jump over Opus 4.8 on benchmarks (not that big vs GPT 5.5 though). Ethan Mollick and Dan Shipper both believe it unlocks new capabilities - a big part of which is its ability to work longer and spawn dozens of subagents reliably without losing context of the main task.

  

  • Fable is only available till 22nd June in the Claude subscription plans. Anthropic plans to move Fable use to paid credits after that until they have the capacity to serve it to everyone. It’s 2x more expensive than Opus (vs Mythos, which was 5x Opus)

  • Anthropic introduced a new policy with the rollout of Fable where it will “secretly” sabotage your work if you use it for any ML/AI-related work. This caused a huge backlash, and in turn Anthropic has partially walked back the “secretly” part of this policy.

  • Other examples of Fable in use - for a refactor to get clean code, for editing a video, and even making a markdown editor.

• Gemini 3.5 Live Translate - new model from Google that does real-time speech-to-speech translation in 70+ languages. Already live in Gemini API and Google Translate + coming soon to Google Meet.

• Implications of large-scale test-time compute - Model performance is affected by how much time, compute, or $$ budget is available for the model to complete a task, and hence companies should report these variables when reporting benchmarks.

• What does AI pricing actually look like in 2026? Orb analyzed 80 AI agent companies, including leaders like GitHub Copilot, Replit, Factory, and Intercom, to identify the pricing models, packaging strategies, and monetization trends emerging across the industry. Download the report.*

Subscribe now

My feed

• Claude Code can have nested subagents now - each subagent can spawn even more subagents. Currently up to a depth of 5 layers.

• The model selector in ChatGPT has been updated to a) show all GPT-5 generation models and b) thinking levels are simplified to Instant, Medium, High, Extra High and Pro - just like how it is in Codex.

• Missions are now available on Factory Desktop.

• Skribe - Local first markdown writing app with an AI review partner.

• Little python harness to run evals on your skills - is the skill improving or hurting the model’s performance?

• pr.video by Mainframe - turn any GitHub PR into a narrated video walkthrough to review changes (without needing the code diff).

• New essay from Dario Amodei on policymaking to keep up with the pace of AI development. It’s a good read but has “trust me bro” vibes at many places.

• Supermemory is now available locally to host yourself.

• DiffusionGemma - new open-weights model from Google that uses a different architecture (diffusion instead of transformers) to gain a 3-5x speedup with roughly similar performance.

• Agents can now sign up on Firecrawl.

Build logs

- by Keshav

I’m building a speech-to-text app that’s entirely local (kinda like Wispr flow, Monologue, Superwhisper, etc.) but no data gets sent to their servers. Part of it stemmed from wanting to play with local models, and part of it was guilt of paying for a similar tool that I don’t use much.

I’m calling the app “Option AFK”, and here’s how I built it in 3-4 days spread out over the past few weeks:

1. Asked Opus 4.7 to write a simple Python script to get Nvidia’s Parakeet 0.6B working on my M3 Air. I tested it in the browser, and the accuracy/speed were better than what I expected.

2. Used Codex’s Computer Use to do a screen-by-screen audit of the tool I was paying for and document all the features that it has with screenshots. I was able to do this on the $20 plan in a single session (hitting the 5-hour limit only once).

3. Then started building the MacOS app with Opus 4.8 (using the audit as a reference). While building this, Opus 4.8 selected an SDK I had no idea existed to wire up things like splitting long voice notes into chunks and speeding the model processing.

Here’s that SDK:

• fluidaudio - run transcription and TTS models on MacOS locally.

I got a working version of the app yesterday with Fable 5, and I’m using it on my device already. Here’s how it looks:

It also supports uploading voice notes (even the longer ones) as a file and getting them transcribed at no additional cost.

I’m planning to release this app once I get the Apple Developer Program sign-up done. Would you want to use it?

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail via @_nicolealonso

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

A lot of chatter about loops on X recently. And it’s a topic I’ve been toying with. My interpretation from what Peter posted is:

Agents are loops, you give it a task, it looks at it with the context it has, uses tools to gather more and gives you an output when it thinks the task is done.

But mimicking those parts into a bigger system so your agents can run more autonomously, longer and on harder chunks of work is what I think ‘designing the loop’ is talking about.

So you want to design a bigger task up front, like a plan.md file with a bunch of tasks, new features to implement, etc. A way for those tasks to be deemed ‘complete, and verified’ - ie. does the report contain all 10 points from the plan, does the UI have all the features working correctly, do all the tests pass. And then prompting itself to go back to the plan.md file and pick up the next one.

I’ve been toying with it for this reference manual - I’m making lots of interactive components, so I’ve tried designing all the component pieces first and then building a workflow to compose those together into the interactive.

But also this is why a ton of people bring skills into their workflows. Do the planning skill, then split tasks with a PRD skill, then research skill for each feature, then building skill, then review skill, then testing skill.

It’s all designing text instructions for an agent to follow, making sure it can access any tools it needs to do the tasks.

Further reading:

• Loop Engineering by Addy Osmani

• WTF is a loop by Matt Van Horn (although AI-sloppy)

Ben’s Bites is brought to you by Smallest AI

Smallest AI Voice Agents gives you production-ready infrastructure to run inbound & outbound campaigns at scale, powered by its in-house realtime latency STT & TTS and enterprise-grade telephony. Enterprises trust it to handle millions of minutes. See what it can do for yours - book a demo!

Headlines

• Apple finally has a dedicated AI product, Siri AI. Imagine about a year-old ChatGPT - with great dictation, image analysis and some interaction with external apps like Messages and Maps. Not bad *if* it works. The new Siri AI uses a mix of local and cloud models (some based on Gemini), all under the AFM 3 model family. These models also power other “AI features” embedded inside apps. I’m keeping an eye out for the one that vibe-codes Safari extensions and Apple Shortcuts using plain English.

• ChatGPT’s memory system runs a background process to save memories that you can see and edit. They are calling the latest iteration Dreaming v3, which has better recall, follows your long-term preferences more closely and corrects itself as time passes.

• New blogpost from Anthropic claims that developers are writing 8x more code (with Claude’s help) than they were in 2025, and it is now helping train the next versions of Claude. Hence, they advocate for an “option” to pause AI development if the need arises.

• OpenAI shared three goals for its next phase: build an automated AI researcher, accelerate the economy and give everyone on Earth a personal AGI. They’ve also filed a confidential S-1 while claiming no urgency for an IPO.

• NotebookLM’s core chat is getting upgraded from the old RAG system to an agent-like system (Antigravity harness). Each notebook gets a cloud computer to run code for analysing the files that you’ve uploaded with the latest Gemini 3.5 models.

• Your Oura ring scores sleep. Your Apple Watch tracks your heart. Workera Ambient does the same for your career: always-on, capability capture from the work that's already happening. Your data, your choice. Learn more from Workera's CEO and join the waitlist.*

Subscribe now

My feed

• Cursor’s Canvas lets users spin up internal apps, dashboards and reports that are shareable with others. Another entry in the “Claude Artifacts but 2026” feature from all the coding agents.

• Can Claude become a chemist? What about powering agents for biology?

• Firecrawl Workflows - installable skills for repeatable web tasks (like deep research, SEO audits and more).

• Eloquent - Local transcription app from Google (uses Gemma).

• FrontierCode - coding eval to test whether code is actually maintainable rather than just passing tests.

• A guide to using /goal in Codex.

• Fin Voice 2 - natural, fast and intelligent customer support over the phone.

• Raindrop 2.0 - catches a production failure, hands it to your coding agent to fix, and turns it into an eval so it can't recur.

• Cognition is guaranteeing up to $10M in credits if Devin underdelivers on an annual enterprise contract.

• skills.sh by Vercel now has an API for querying its collection of 600k+ skills.

• Upstash Agent Analytics - 3 lines of code to track AI/agent traffic to your website.

• Spiral by Every - writing partner for humans and agents with stylometry, CLI, MCP/API, team styles.

• Google is making its budget AI plan even cheaper ($7.99/mo to $4.99/mo) while offering 2x the storage space.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com