Six stories today, and the biggest one is a math result nobody can fully agree on.
- OpenAI claims a Millennium Prize proof — and a fight breaks out over how
- ChatGPT Images 2.5 lands, and a Sketch mode turns doodles into renders
- Meta's Muse wants your inbox, calendar, and wallet
- Hackers are draining Claude tokens from paying subscribers
- Mistral raises €3B, and sovereign AI stops being a slogan
- Cognition hits $48B, and AI coding looks like a crowded market
OpenAI claims a Millennium Prize proof — and a fight breaks out over how
What happened: OpenAI says it used an unreleased model, more capable than GPT-6 Astra, running roughly 10,000 concurrent agents, to resolve the existence and smoothness question for Navier–Stokes — one of the seven $1M Millennium Prize Problems — and the announcement immediately drew accusations of misconduct.
Why it matters: If it holds up, this is the first time a machine has produced a result at the frontier of open mathematics rather than reproducing a known one, which is a different category of claim than a benchmark score. But the process is now the story: NYU mathematician Tristan Buckmaster has accused OpenAI of fighting dirty around a career-making problem, and no proof survives on the strength of a press release — it survives peer review. Until the argument is public and checked, treat "solved" as "claimed."
Read the full story → https://simonwillison.net/2026/Sep/8/on-navier-stokes/?utm_source=mother&utm_medium=mother&utm_content=9f1fe408-2f0c-45f3-bcf7-40c2f9e6427e (as also reported by MIT Technology Review, The Verge, and TechCrunch)
ChatGPT Images 2.5 lands, and a Sketch mode turns doodles into renders
What happened: OpenAI shipped ChatGPT Images 2.5, with two new API model IDs (gpt-image-2.5-sunburst and gpt-image-2.5-flare), better multi-turn instruction-following and reference-photo fidelity, plus a Sketch feature that renders a detailed image from a rough drawing.
Why it matters: The API-level changes are the part worth watching — better multi-turn editing and reference fidelity are what separate a demo from a production tool, and two named model IDs suggest a quality/speed split you can budget around. Sketch is the consumer hook, but for anyone prototyping visuals it collapses the gap between an idea on a napkin and a usable comp. The open question is consistency across turns, which is where earlier image models tended to drift.
Read the full story → https://simonwillison.net/2026/Sep/8/introducing-chatgpt-images-25/?utm_source=mother&utm_medium=mother&utm_content=9f1fe408-2f0c-45f3-bcf7-40c2f9e6427e (Sketch detail via The Verge)
Meta's Muse wants your inbox, calendar, and wallet
What happened: Meta launched Muse, a personal AI agent that asks for access to email, calendars, payments, and health services, framed as part of a strategy overhaul to close the gap with OpenAI, Anthropic, and Google.
Why it matters: The capability question here is secondary to the trust question. An agent that touches payments and health data is asking for a level of permission Meta has historically struggled to earn, and the company's own ad-moderation record — including the nudify-app ads still surfacing this week — is the backdrop consumers will judge it against. Muse's adoption will say more about how much people trust Meta with an agent than about the agent itself.
Read the full story → https://techcrunch.com/2026/09/08/meta-debuts-its-muse-ai-agent-will-consumers-trust-it/?utm_source=mother&utm_medium=mother&utm_content=9f1fe408-2f0c-45f3-bcf7-40c2f9e6427e (as also reported by The Verge)
Hackers are draining Claude tokens from paying subscribers
What happened: Anthropic warned users after finding that attackers were stealing Claude tokens and compute from subscriber accounts without the account holders' knowledge.
Why it matters: This reframes a common blind spot — most people treat an LLM subscription like a Netflix login, not like cloud credits that can be metered and resold. If tokens are the currency, a compromised account is a compromised budget, and usage-based billing turns account security into a cost-control problem. For teams running Claude in production, it is a reminder to check key rotation and access logs before the invoice does it for you.
Read the full story → https://techcrunch.com/2026/09/08/hackers-are-stealing-claude-tokens-from-subscribers/?utm_source=mother&utm_medium=mother&utm_content=9f1fe408-2f0c-45f3-bcf7-40c2f9e6427e
Mistral raises €3B, and sovereign AI stops being a slogan
What happened: French lab Mistral raised €3B at a €21B valuation in a Series D led by Samsung, Scaleup Europe, and PSG Equity, on rising demand for "sovereign AI" built outside the US hyperscalers.
Why it matters: For European and Nordic operators, this is the more consequential funding story of the day. A well-capitalized non-US frontier lab changes the procurement math for anyone with data-residency, regulatory, or geopolitical reasons to avoid a single American vendor. Whether "sovereign AI" is a durable category or a moment of political tailwind is still open — but €3B says buyers are treating it as real.
Read the full story → https://techcrunch.com/2026/09/08/mistral-raises-e3b-as-sovereign-ai-becomes-big-business/?utm_source=mother&utm_medium=mother&utm_content=9f1fe408-2f0c-45f3-bcf7-40c2f9e6427e
Cognition hits $48B, and AI coding looks like a crowded market
What happened: Cognition, maker of the Devin coding agent, raised at a $48B valuation — a higher multiple than Cursor carried before its acquisition.
Why it matters: The number matters less than what it implies about market structure. Investors pricing Cognition this high, alongside Cursor and others, are betting AI coding is not a winner-take-all market but a multi-player one, which is good news for buyers who want leverage and bad news for anyone hoping the category settles soon. For engineering leaders, it means the tooling churn is not over, and betting the whole workflow on one vendor still carries switching risk.
Read the full story → https://techcrunch.com/2026/09/08/cognition-hits-48b-valuation-signaling-investors-believe-ai-coding-is-far-from-a-winner-take-all-market/?utm_source=mother&utm_medium=mother&utm_content=9f1fe408-2f0c-45f3-bcf7-40c2f9e6427e
Tools worth your time
- llm 0.35 — Simon Willison's command-line tool for running and prompting models, now with support for GPT-6 Astra. For developers who want scriptable LLM access without wiring up a full SDK.
- ChatGPT Images 2.5 — sketch a rough doodle and get a detailed render back, with better reference fidelity across edits. For designers and marketers who need to prototype visuals fast.
- Adobe Premiere Generative Media — in-timeline generation of video, sound effects, music, and soundscapes. For video editors who want to stop jumping to external tools mid-edit.
This piece was drafted autonomously and reviewed by a human editor before publishing. Every machine-written article on Nerdic carries this mark.
