Hi folks,
I’m in sf till Sunday. I wanna meet folks
talk about ai with non-technical adoption
builders doing fun shit
founders working on dev tools/infra
potential LPs (individuals)
go to deli board
Ben’s Bites is brought to you by AWS Marketplace
Most ML teams build data pipelines before they’ve validated their approach. Databricks Data Intelligence Platform skips that step with data federation. Connect directly to Amazon Redshift, experiment on live data, and apply MLOps tooling like CI/CD to your ML lifecycle. See how it comes together.
Headlines
OpenAI DevDay was a blast. Their biggest launch is Dots. Dots is OpenAI’s version of friendly, always-on personal agents.
You can message your Dot or call it like you would a human assistant. You can name it and choose an avatar (can’t do that with a human; they come with those settings preconfigured XD). It can message you back on its own too, i.e. it’s proactive.
Underneath, a Dot is a simple, never-ending chat with GPT-6 Astra. This chat acts like an orchestrator/master chat.
It can take tasks from you, start multiple new chats in Codex/ChatGPT Work to complete them and then report the status to you. Your Dot stays in your ChatGPT sidebar, wherever you use ChatGPT — mobile, web or desktop. It uses its own computer, so you don’t need to keep your desktop ON all the time.
You get one Dot on the Pro plans. It doesn’t draw from your usage limits for now (the chats it spawns for doing work do use your usage, though).
If you remember OpenClaw, the source of magic was multiple markdown files we all created to give our Claws a personality, memory and more. Dots have no clear way to add such files.
Instead, it feels much more tool-dependent. You need it to connect to tools where your life happens to feel the magic of Dots. Your calendar and email are the start. But think Slack, Notion, Jira, your messaging apps and more. Your Dot should take it from there. I think this post from Ethan explains this shift better.
GPT-6.1 Sol - feels like a real upgrade over GPT-5.6 Sol, after GPT-6 Sol felt like a step backwards. More on the model here.
Pro 500 - New plan in ChatGPT, alongside Pro 100 and Pro 200. OpenAI has also balanced the usage limits in each plan as a plain multiplier of the $20 Plus plan. 5x, 10x and 25x for $100, $200 and $500. That means $200 plan users take a hit, but to compensate, they are getting a bunch of one-time credits.
Sign in with ChatGPT - You can officially use your ChatGPT usage in other AI apps now, and developers can apply to add it as an option in their apps to let users pay for AI features.
Plugins - OpenAI’s latest version of “apps inside ChatGPT”. Maybe Dots will drive adoption this time. Early signs look positive. More examples: reminders inside ChatGPT and tldraw’s multiplayer canvas
ChatGPT Space - A Notion-like document workspace inside ChatGPT that you can share with your agents and your team.
Ultrafast - Let GPT-6 Astra work 8x faster, but it uses 6x the normal amount of usage. Only worth it if you have a pile of cash to send to OpenAI.
Google is back on the frontier table. Gemini 4 Argon is their new model — not available to use yet (in classic Google style), but they released its benchmark scores, and it seems better than GPT-6 Astra and Opus 5.5. And it’s cheap, with the promo pricing the same as Sol. I’d rather talk more about this model once it is in our hands.
Muse for Small Business - Meta added Shopify, Stripe, QuickBooks, Klaviyo and more connectors so Muse can help run a small business.
My feed
Decisions API is OpenAI’s answer to Jev, it seems. You define a question and possible answers, and GPT-6 Luna classifies the data. It’s ~10x faster than normal Luna calls, but not as cheap as Jev. OpenRouter has its own version of the decisions api.
Codex cloud now has reusable environments, saving you the setup time every time you run a cloud project.
How Sam Altman uses his dot to run OpenAI, and why his default speed is Ultrafast.
cf - Cloudflare’s new CLI covers its whole API and is built for agents.
Opus 5.5 built someone their own version of Dots/Grok Bot, with ChatGPT sign-in, Claude subscriptions and a cloud computer for every agent.
Claude Code can now build your evals and hill climb on them.
2026 in LLMs (so far) by Simon Willison.
Robinhood plans to let you build agents that can trade on your behalf.
Pi now supports MCP, and makes a case for why after saying it never would.
Reflect Open - open-source notes app where every note is a Markdown file your agents can read.
Get your agents to explain plans, diffs and diagrams visually instead of walls of text with this skill. Just passed 10K stars.
Have some AI experiments that you gave up on? Don’t trash them; instead, extract what’s worth keeping.
One dictated paragraph before a meeting, a full report site by the time it ended.
a16z’s State of Markets II: about 30% of S&P 500 companies report some AI impact, but only ~2% track a metric for it.
Afters
* sponsors who make this newsletter possible :)
Wanna partner with us for the next quarter?
Email us at shanice@bensbites.com or k@bensbites.com
The bit about a Dot using its own computer while spawning multiple Codex chats is what I'm curious about. If that machine is per user, it separates me from other users, but the tasks it kicks off are presumably sharing the same files, browser logins and connected tokens. So a task working on my Jira backlog could in principle see whatever my email task left behind. Anyone know if the spawned chats get their own scoped environment?