The Brief #26.37 · GPT-6 Astra, AI math proofs, and NVIDIA buys Hugging Face

Artificial Analysis swapped the tests behind the model table so every score moved, Chrome 153 shipped camera and microphone as real HTML elements, OpenAI launched GPT-6 Astra and published an AI-written Navier-Stokes proof, Claude, ChatGPT, and Grok broke on the same morning, and NVIDIA signed to buy Hugging Face.

The model table was rebuilt, and every score fell

Artificial Analysis replaced the tests behind its Intelligence Index twice this week, releasing v4.2 on September 4 and v4.3 on September 7. The index is the number this table reports every week, an average of the same fixed set of hard tests run on every model. Last week that set was nine tests. The new set is ten, with bigger and longer tasks.

GPQA Diamond is out because top models now ace it. Held-out tests (questions or answers that never get published, so a lab can’t train its model on them) now supply 45% of the score.

Every score fell, most by a fifth to a third, and cost per task at least doubled. No model got worse, and the posted token prices didn’t move. The tasks are simply larger, so the same model spends more tokens finishing one at the same token prices. Claude Fable 5.1 printed 66 in last week’s Brief and reads 53 today, and it’s the same model at the same posted price.

The two revisions landed in the week GPT-6 Astra launched, and readers on Hacker News noticed the timing. Artificial Analysis says the updates are pieces of its planned v5 arriving early “to keep pace with the frontier”, after it deliberately held them back through recent launches.

At the top, Fable 5.1 and Astra both display 53, and Fable stays ahead by half a point underneath, 53.37 against 52.81. Astra gets its 53 for $3.26 a task against $7.63 for Fable, so same performance for half the price.

Table The Intelligence Index top fifteen on the new v4.3 test set, plus Claude Fable 5.1 at the effort Claude Code uses.

#	Model	Intelligence Index v4.3	Cost per Index task
1	Claude Fable 5.1 (max)	53	$7.63
2	GPT-6 Astra (max)New	53	$3.26
3	Claude Fable 5.1 (high)	51	$3.91
4	Claude Opus 5 (max)	51	$5.86
5	Claude Fable 5 (max)	50	$8.75
6	Muse Spark 1.3 (max)New	48	$1.60
7	GPT-5.6 Sol (max)	47	$1.99
8	GLM-5.3 (max)	45	$2.01
9	Grok 4.6 (high)	44	$1.86
10	Kimi K3 (max)	44	$2.00
11	GPT-5.6 Terra (max)	42	$1.40
12	GLM-5.3-Flash	42	$0.25
13	Gemini 3.8 Flash (high)	41	$1.24
14	Qwen3.8 Max	40	$2.67
15	Qwen3.8 2.4T A95BNew	40	$2.16
16	Muse Spark 1.2 (xhigh)	40	$0.97

Chrome 153 makes camera and microphone real HTML elements

Chrome 153 started rolling out on September 8. It adds <camera> and <microphone>, HTML elements the browser itself draws and controls, so a page can’t restyle them into something misleading. The permission prompt only appears after a real click on that browser-drawn control. They’re the third and fourth of these capability elements, after <geolocation> in Chrome 144 and <usermedia> in 151.

CSS gets scroll-axis-lock, which keeps the browser from locking a scroll gesture to one axis, and JavaScript gets Iterator.zip and a join() method on iterators.

Separately, the September 3 Chrome 152 security update fixed CVE-2026-85046, a bug in V8, the engine that runs JavaScript inside the browser, which Google says has an exploit in the wild. NVD, the U.S. vulnerability database, describes it as code execution inside the sandbox rather than a sandbox escape. Restart your browser if it’s been open a while.

OpenAI launched GPT-6 Astra

OpenAI launched GPT-6 Astra on September 3, and it’s available in ChatGPT Plus, Pro, Business, and Enterprise, the API, Microsoft Azure, AWS Bedrock, and GitHub Copilot. In Copilot it’s on the Pro+, Max, Business, and Enterprise plans, billed at OpenAI’s list price on top of the subscription.

The API price is $10 input / $50 output per million tokens, 2.5 times GPT-5.6 Sol, for a context window just over a million tokens. The model always reasons (the API accepts no none effort setting), and tool calls (one round trip where the model asks to read a file or run a command) only work through the Responses API. It’s also the first model OpenAI rates Critical for cybersecurity, the level where it can find and exploit unknown security flaws on its own.

OpenAI’s launch page says Astra “saturates ARC-AGI-3 with a 99.9% score”, as in it maxes out the test. The benchmark’s own maintainers at ARC Prize measured 62.7% the same day by running Astra the way they run every model. The 99.9% shows up only in a special setup that lets Astra keep its private reasoning between requests. ARC Prize also wrote that it’s “not claiming that it is AGI”.

Fermat’s Last Theorem, then Navier-Stokes

Anthropic published the first complete machine-checked proof of Fermat’s Last Theorem on September 4. The theorem says the equation a^n + b^n = c^n has no solution in positive whole numbers once n is bigger than 2. Mathematicians searched for a proof for more than 350 years, and the one Andrew Wiles finally published in 1995 ran 129 pages and took months of human checking. Claude wrote this one largely on its own over 11 days in Lean, a language where a computer checks every step, so nobody has to take anyone’s word for it.

The same week, OpenAI went after one of the biggest open problems in math. The Clay Mathematics Institute keeps a list of seven famous unsolved problems and offers a million-dollar prize to whoever solves one first. One of them asks whether the Navier-Stokes equations, the math behind how fluids move, allow a smooth fluid to reach infinite speed in a finite time. OpenAI’s post says a rumor reached the company on September 1 that someone had already solved two of those problems. No proof was public anywhere, so OpenAI set about 10,000 agents on the fluid question at once to produce one of its own. Seven days later, on September 8, it published a 166-page proof, plus a Lean version, claiming that with a smooth outside push a fluid really can blow up to infinite speed while its energy stays finite. That counts as one official version of the problem, but the famous one, where the fluid blows up with nothing pushing it, is still open.

An unnamed internal model the company calls “significantly more capable than GPT-6 Astra” produced the proof, with the agents running for about 88 hours and Astra writing the final Lean version in 17 hours. The same day, mathematicians Levent Alpoge and Tristan Buckmaster released their own proofs of the same kind of blowup for three related fluid equations, two of them machine-checked, after months of work. Terence Tao called it “A remarkable achievement” and wrote that the pair “had to release far earlier than planned due to external events”. A result of the same kind took the mathematicians months and the agent fleet under four days.

One detail makes the timing awkward. Buckmaster’s statement says the pair had put every draft through Codex, OpenAI’s coding tool, for the whole project. OpenAI’s post says “no specific user data was accessed”, while also saying “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models”. Buckmaster writes “I am not accusing anyone of anything.” So, it seems it wasn’t a “rumor” after all.

NVIDIA signed the Hugging Face deal

Last week this was a report neither company would confirm. On September 3, Jensen Huang announced that NVIDIA has agreed to buy Hugging Face, the place most open models and datasets get published and downloaded, for $12,930,300,000.

The securities filing behind the announcement says the agreement was signed on September 2. It splits the sum into about $11.9 billion for Hugging Face stockholders plus up to about $1 billion in retention pay for employees joining NVIDIA. The close is expected in the first half of 2027, pending regulatory approval.

The filing commits NVIDIA to keeping the platform open, with anyone free to upload and download the models they choose and other chip makers still supported. Huang’s post adds that “NVIDIA compute will not be required to build on or deploy through Hugging Face”. Neither document says anything about pricing, so nothing in writing keeps today’s free tier free.

The number itself is a deliberate joke, since $12,930,300,000 starts with 129303, the hugging face emoji’s Unicode number in decimal, and that emoji is the company’s logo.

ChatGPT Images 2.5 draws from your sketches

OpenAI released ChatGPT Images 2.5 on September 8, rolling out to all ChatGPT tiers. The chat side gets Sketch, where you draw directly in ChatGPT and the model treats the drawing as a reference (type @Sketch to start). Templates cover common formats like posters, comments can now sit directly on images, and a shared image can carry its prompt. The company says people create more than 3 billion images a week across ChatGPT and the API, and that it cut generation latency by up to 50% from Images 2.0.

The API side gets two new models. OpenAI’s default pick, gpt-image-2.5-flare, makes higher-quality images than GPT Image 2 at 50% lower latency by the company’s measure, while gpt-image-2.5-sunburst trades speed for tighter control across edits. Both add xhigh and max quality settings, and token prices match GPT Image 2, $8 image input / $30 image output per million tokens, with text input at $5.

Muse Spark 1.3 enters the table cheap and fast

Meta released Muse Spark 1.3 on September 2, available in Muse Code and the Meta Model API. It costs $1.25 input / $4.25 output per million tokens, enters this week’s table at 48, and at 232 tokens a second it moves faster than every model above it on the table. In Meta’s own comparisons it makes about 20% fewer tool calls and uses about 25% fewer tokens than 1.2. Meta lists an open-weights release (the model files free to download) on its roadmap, and until that lands the model runs only on Meta’s services.

Interop 2027 takes proposals until September 23

Interop 2027 opened for proposals on September 3, with a September 23 deadline. Interop is the yearly agreement where the browser makers pick a set of cross-browser problems and all fix them the same way, scored in public by shared tests. If a feature keeps working differently across browsers for you, file it before the deadline, because next year’s shared fix list gets picked from these proposals. In the first five days people filed more than 40 proposals, with Temporal, CSS line-clamp, and anchor-position container queries among them.