The newsletter for ai builders of all levels. Mini-tutorials, tool reviews, and lay of the land from an exited founder turned investor and forever tinkerer.
Hey folks,
GPT-5 is live and not everyone likes it. OpenAI reduced the model picker to two options: GPT-5 and GPT-5 Thinking. All the older models—4o, o3, o4-mini etc. are gone from the app.
GPT-5 is still a collection of models behind the scenes. Primarily 4 models;
GPT-5 and GPT-5 mini (supposed replacements for GPT-4o and GPT-4o-mini)
GPT-5 thinking and GPT-5 thinking mini (replacements for o3 and o4-mini)
The router picks the best model for your query. But it broke on day 1 😬.
They’re all better than what existed in ChatGPT before, but the performance is in line with the general rate of progress across all other models— ie, there’s no breakthrough. It has similar performance to Claude Opus 4.1 and Gemini 2.5 Pro.
There was some pushback on removing 4o, because people liked its personality. So even though newer models have less hallucinations and are more powerful - they miss the ‘vibes’.
Every time there’s a new model it just takes some time to get to grips with.
This is the general sentiment I agree with:
It's my new favourite model - it's not a giant leap forward, but it's noticeably competent at everything I've thrown at it. — Simon Willison.
You can go back to using 4o by enabling “legacy models” in ChatGPT and more visibility into what model is being used is coming to the app.
ICYMI, we published a post last week of what tinkering with AI and code looks like for us.
Attio is the AI-native CRM for the next generation of teams. Sync your email and calendar, and Attio instantly builds your CRM—enriching every company, contact, and interaction with actionable insights in seconds. Join fast growing teams like Granola, Flatfile, Modal, and more. Start for free today.*
From the portfolio:
Anything - a single place to build (rebranded from ‘Create’).
Rork adds an agent to build mobile apps.
GPT-5 is the default model in Factory AI’s droids.
Pika built a new audio-driven model. It generates HD videos that match the energy and expressions to the audio, cheaper and faster than before.
Folks at Gamma made this cute cat game to collect your OOO hats.
*sponsored
🌐 What I’m consuming
How much do we wanna defer our decision-making to AIs?
Why all scores on SWE-bench Verified are not the same.
Using Google Stitch and Jules to design and build an app.
GPT-5 review by Latent Space
GPT-5 vibe check by Every
⚙️ New tools and updates
Graphy - From messy data to beautiful graphs in a click.
Meteor by browse.dev - An intelligent browser on autopilot to understand pages, suggest actions, and automate tasks with one click.
v0 by vercel gets a tiny rebrand moving from v0.dev → v0.app and a big upgrade with agentic planning, research and debugging tools for building software.
Claude can now reference past chats. available in Max ($200/mo) today and coming to Pro ($20/mo) soon.
Grok 4 is now available for free users to try, and it can also read PDFs with 100s of pages.
*sponsored
🥣 Dev dish
In the API, all GPT‑5 models can accept a maximum of 272,000 input tokens and emit a maximum of 128,000 reasoning & output tokens, for a total context length of 400,000 tokens. OpenAI API now also supports custom tools written in plain text so you’re not limited to JSON.
Cursor is also jumping into your terminal with a CLI agent.
AI Elements by AISDK - Beautifully designed React components for building AI interfaces.
Vibe coding agent template by Vercel.
Claude Code can now handle long-running tasks in the background.
🍦 Afters
Anthropic’s Red Team has a new blog covering their internal research on cyber, bio, autonomy, national security and more.
Leopold Aschenbrenner, a former OpenAI employee, wrote a viral essay titled “Situational Awareness” last June. In an year, he has built a hedge fund with $1.5B within an year.
That’s it for today. Feel free to comment and share your thoughts. 👋
📷 thumbnail creds: @keshavatearth
GPT-5 just landed and the hot takes are already stacking up—no giant leap, just steady polish. While everyone debates vibes vs. benchmarks, keep your own testing loop tight: open https://digitalclock.live/ in fullscreen and time how long each prompt actually takes to answer; the live second hand turns every trial into hard data instead of gut feel.