Hello AI Enthusiast,
Three big releases this week, and all of them land the same way, with the demo running ahead of the daily work. OpenAI says an internal model that nobody outside the company can touch has proved a problem mathematicians have chased for decades, and the fight over who deserves the credit is still going. GPT-6 Astra tops the benchmarks, though most teams we work with are well served by something smaller. And Meta's new agent will buy your groceries, as long as your shop is one it supports. Let's dive in.
The Big Picture 🔊
OpenAI Says an Internal Model Cracked a Millennium Prize Problem
OpenAI published a proof that a smooth fluid at rest can develop a singularity in finite time, which resolves statements C and D of the Navier-Stokes Millennium Prize formulation. The work came from an internal system that OpenAI says is more capable than GPT-6 Astra, running about 10,000 agents in parallel, 2.7 million messages and roughly 130 billion output tokens across 88 hours.
|
||
|
Two things caught our attention. The proof was checked by software and has not been signed off by the Clay Institute, so the headline is running ahead of the verdict. And the system behind it is internal, which tells us more about what OpenAI has in the building than about the tools any of us can use next Monday. The credit fight with the researchers who were working on the neighboring problem is worth following too, because it shows how these labs behave when there is a trophy on the table. |
GPT-6 Astra Arrives, and It Is Very Good at Using a Computer
OpenAI released GPT-6 Astra, its new frontier model. The numbers it puts forward are computer use, 72.6% on OSWorld 2.0 and 47% faster than GPT-5.6 Sol, plus 95.9% on 3D object reconstruction and 98% on FrontierMath Tier 4. It goes to selected organizations first, then paid ChatGPT tiers, the API, Azure and Bedrock, at $10 per million input tokens.
|
||
|
The benchmarks are strong and the demos are stronger, so it is worth separating the two. In daily work most of us are already well served by mid-range models, and we feel the difference mainly on visual and very long tasks, like building a 45 slide deck or reading through a folder of contracts in one go. If your team is not hitting a wall with the model they have today, a smarter model is not the upgrade that changes their output. |
Meta Launches Muse, an Agent That Actually Buys Things
Meta introduced Muse, a personal AI agent that acts instead of only chatting. It turns saved Instagram recipes into grocery lists, books travel, sends emails and checks out through Link by Stripe, and it keeps working when the app is closed. Muse lives in its own iOS and Android app, inside WhatsApp, and soon on Meta's glasses. United States only for now.
|
||
|
Calling it the world's first personal AI agent is a bold line for a product arriving this late. The part we find interesting is distribution, because Meta can put an agent in front of people who already spend hours a day inside its apps, and that beats asking anyone to open something new. The limit is the one every agent runs into: it works beautifully when your shops, your bank and your tools are the ones it supports, and falls apart when they are not. |
Bits and Bobs 🗞️
NVIDIA is acquiring Hugging Face, and says the platform will stay open and multi-cloud.
OpenAI released Images 2.5, with sharper detail, faster generation and better handling of reference photos.
ChatGPT Work can now learn your writing style from the tools you connect, including Gmail, Drive, Slack and SharePoint.
Microsoft opened up Project Opal in Copilot, which runs multi-step browser tasks on Windows 365 Cloud PCs.
Google added agentic video to Gemini, cutting token use on long videos by up to 88%.
Google also shipped Gemini 3.8 Flash and 3.8 Flash Cyber, which finds vulnerabilities across 20 programming languages.
Lyria 3.5 in the Gemini app now makes custom jingles and ringtones, in case you needed one.
Anthropic published a blueprint for commerce agents so retailers can build shopping agents on Claude.
OpenAI committed $1 billion to Daybreak, a cybersecurity program for organizations running essential public services.
OpenAI is also putting $5 million into research on how generative AI affects teenagers aged 13 to 17.
OpenAI shared how coding agents are speeding up its own research, with humans still supervising.
From Our Founder’s Channels 🤳
In his latest LinkedIn video post Gianluca Mauro talks about the AI effect: people's expectation of what is "AI" shifts as AI itself evolves. Watch it here.
A new model drops almost every week, and it's hard to tell what actually matters for your team. Most of it sounds impressive but won't change how your team works day to day. Our Corporate Training is about turning all that noise into real, everyday results.
That's a wrap on our newsletter!
If you're looking to build AI skills in your team, check out our Corporate Training. Every program is tailored to your team and built around hands-on practice.
Catch you next week! 👋
