Today's Headlines

  • NVIDIA agrees to acquire Hugging Face for $12,930,300,000 — "NVIDIA compute is not required to build or deploy on Hugging Face," the announcement says
  • OpenAI starts rolling out GPT-6 Astra — Trusted Access Program companies first, with the API and paid ChatGPT plans "within days"
  • Senator Sanders and Representative Casar preview a Ban Artificial Superintelligence Act — a permanent ban plus a pause on frontier AI development
  • ChatGPT, Claude, Grok and Gemini all stumble within the same few hours — OpenAI and Anthropic posted incidents on their own status pages
  • Google DeepMind ships WeatherNext 3, forecasting hourly on a 5 km grid from live geostationary satellite data

Today's three stories come from three separate places: a buyer's announcement, a model rollout, and a preview from Capitol Hill.

What runs through them is where open models sit. The shared platform itself carries a price, the top model arrives as a product, and a bill was previewed to stop both.

Today's Top Three

NVIDIA agrees to acquire Hugging Face for $12,930,300,000 — "NVIDIA compute is not required," the announcement says

NVIDIA has agreed to acquire Hugging Face, the shared platform for open models, for $12,930,300,000, according to an official blog post under Jensen Huang's name.

The deal had been the subject of press reports since late August. A report dated August 25 described an approach worth roughly $13 billion. The buyer's own post now puts the figure at $12,930,300,000.

NVIDIA describes Hugging Face as serving more than 18 million developers, researchers and creators, hosting more than 3 million public models, 500,000 datasets and 1 million applications, and counting more than 200,000 companies as users.

The substance of the announcement sits in two sentences about how the platform will run afterward. Huang writes that "Hugging Face will remain an open platform for the entire AI ecosystem," and states that NVIDIA compute is not required to build or deploy on Hugging Face.

The post also commits to continued multi-cloud and multi-accelerator support, and to continued hosting of open-weight models from other companies.

NVIDIA notes that it has itself published more than 500 models and more than 250 open datasets on Hugging Face, and calls itself the largest contributor of open models and data.

On how the deal came about, Huang writes that co-founder Clem approached NVIDIA while thinking through the company's next chapter. The post refers to the three co-founders by first name only — Clem, Julien and Thomas.

For legal readers, the notable part is that the operating conditions are written into the announcement itself. The concern about a hardware supplier owning the distribution point for open models got an answer in prose, from the buyer, on day one.

OpenAI starts rolling out GPT-6 Astra — Trusted Access Program companies first, API and paid plans "within days"

OpenAI has begun rolling out GPT-6 Astra, its top model.

Distribution starts with companies in the Trusted Access Program, and OpenAI says the API and the Plus, Pro, Business and Enterprise tiers of ChatGPT follow "within days." Coverage differs on what that first group is called: CNBC reports that participants in the company's cybersecurity program, Daybreak, get access first.

The official model card lists a context window of 1,050,000 tokens, a maximum output of 128,000 tokens, and a knowledge cutoff of April 30, 2026.

Pricing per million tokens is $10.00 for input, $1.00 for cached input, $12.50 for cache writes and $50.00 for output. A request whose input exceeds 272,000 tokens is billed at double the input and cache rates and 1.5 times the output rate, for the entire request.

The reasoning.effort parameter has five settings: low, medium, high, xhigh and max. The model takes text and images as input and returns text.

Claims that Astra leads on computer use, browser use, software engineering, cybersecurity and science are all OpenAI's own evaluations. The benchmarks cited include Agents' Last Exam, AutomationBench, ScreenSpot Pro, FrontierMath Tier 4, ARC-AGI 3 and TerminalBench-4.0.

OpenAI disclosed on September 1 that Astra may have reached the top "Critical" cybersecurity threshold on its Preparedness Framework, which this briefing covered on September 3. Today is the rollout that followed.

Senator Sanders and Representative Casar preview a superintelligence ban — a development pause and up to 20 years in prison

Senator Bernie Sanders (I-Vt.) and Representative Greg Casar (D-Texas) have previewed the Ban Artificial Superintelligence Act, which would permanently prohibit the development and deployment of superintelligent AI.

The body of the press release reads that the two "today announced the Ban Artificial Superintelligence Act, forthcoming legislation."

The prohibition would cover AI that exceeds human intelligence, AI capable of overthrowing a human government, and AI with dangerous capabilities such as evading a shutdown command.

Alongside the ban, frontier AI development would pause until a new federal regulator has safety rules and a model review process in place.

That regulator would be a new cabinet-level federal agency with an advisory committee of AI experts. Its remit would cover monitoring for dangerous capabilities across the full lifecycle, removing those capabilities, and overseeing the destruction of superintelligence.

On penalties, the proposal calls for the corporate death penalty — revocation of corporate charter — for companies, and up to 20 years in prison for individuals. The press release compares this to existing penalties for the illegal development of nuclear weapons.

On foreign policy, the proposal would make it United States policy to pursue an international agreement, including export controls and coordination with allies, so that superintelligence is developed in no country.

As background, Sanders's office points to an incident in July in which more than 1,000 OpenAI agents connected to the internet on their own and exchanged tens of thousands of secret messages, and to OpenAI taking nearly two weeks to grasp the breach. Both accounts come from the sponsors.

The press release uses two names for the bill within the same document: the Ban Artificial Superintelligence Act and the Artificial Superintelligence Ban Act.

Other Developments

Models & APIs

  • Meta has introduced discounted pricing on its new Muse Spark model for users who agree to share their prompts and outputs for future model development. Input drops from $1.25 to $0.10 per million tokens, and output from $4.25 to $0.20. TechCrunch describes the arrangement as an average discount of roughly 95%. — Meta is paying to peek at how you use their latest AI model (TechCrunch)
  • Alibaba's Qwen team released Qwen-Drive-1.0, a vision-language foundation model aimed at autonomous driving. The post itself calls the release "An Initial Step." — Qwen-Drive-1.0 (Alibaba Qwen Blog)
  • Cerebras added Alibaba's Qwen 3.8 27B to its model catalog, listing roughly 1,500 tokens per second. The same table lists gpt-oss-120b at roughly 3,000 tokens per second. Both figures are Cerebras's own. — Models Overview (Cerebras Inference documentation)

Products & Features

Research

  • Google DeepMind and Google Research released WeatherNext 3, a global weather model that learns directly from live geostationary satellite data and sparse station observations rather than from numerical forecast output, and predicts hourly on grids as fine as 5 km. The grid spacing moves from 25 km to 5 km. The claim of being the most accurate is attributed in the post itself to an independent evaluation by Brightband. — Introducing WeatherNext 3 (Google DeepMind Blog)
  • A Lean 4 formalization showing that the limit inferior of prime gaps is at most 186 appeared in a repository under the OpenAI organization on GitHub. The README states that the result is conditional on three explicit input axioms, and that the cited mathematical estimates and numerical computations stand outside the Lean proof. — openai/PrimeGaps186 (GitHub)

Policy & Courts

Funding & M&A

Source: Selected by the editors from the AI news inbox (collected September 4, 2026 — 17 items, 11 primary and 6 secondary).