OpenAI pauses Astra over cyber risk + a $16.8 billion Musk chip fab

OpenAI disclosed that internal evaluations of Astra, a model still in development, show enough strength in agentic coding and cybersecurity that the company "cannot rule out Critical capability level at this time" — the first time that tier of its Preparedness Framework has been triggered. The framework defines Critical cyber capability as a model that can find and build working zero-day exploits in hardened real-world systems without human help, or run end-to-end attacks on hardened targets given only a goal. OpenAI paused internal Astra work that lacks the new safeguards, turned on universal monitoring of the model, and is bringing in government agencies and outside safety organizations to test it. The company added that Astra was not the model involved in the recent Hugging Face breach.

The two companies confirmed Grimes County, Texas — about an hour northwest of Houston — as the site for Terafab, a plant that would run logic, memory, packaging and testing under one roof rather than splitting them across suppliers. The first phase is $16.8 billion for a facility of more than 100 million square feet and at least 3,000 workers, making chips for Tesla's Optimus robots and Cybercab as well as SpaceX's planned space-based data centers. The caveat is in SpaceX's own filing, which describes the arrangement as a general framework with no binding commitments and no obligation for either side to keep going. No completion date has been given.

The Financial Times reports that ByteDance has a model of roughly 10 trillion parameters in pre-training, a stage that usually runs three to six months before fine-tuning and release. That would be more than three times the size of Moonshot's Kimi K3 at 2.8 trillion parameters, currently the largest Chinese model with published weights, and well past DeepSeek's V4-Pro and Meituan's LongCat-2.0 at about 1.6 trillion each. The Mythos comparison is softer than it looks: Anthropic does not publish parameter counts, so the roughly 8 trillion figure attached to Mythos 5 is an industry estimate, not a disclosed number. Reuters said it could not verify the report, and ByteDance did not comment.

Kitesurf drops the parts of a browser a human needs — tabs, themes, extensions, visual chrome — and keeps what an agent needs, running entirely on Cloudflare Workers instead of a Chromium process. It is assembled from Rust pieces: the Blitz rendering engine, Firefox's Stylo CSS parser, Boa for JavaScript and Parley for text. Cloudflare measures 3 to 4 times less CPU and 5 to 7 times less memory than Chromium on screenshot and HTML-extraction jobs, at the cost of being a little under twice as slow in wall-clock time. It is free during beta through Browser Run, and Cloudflare says it plans to open-source the code once it is ready.

Rippling was on track to spend 40 percent of its research-and-development headcount budget on AI tokens, with spend growing 80 percent month over month from early 2026 and one engineer alone running $50,000 a month. Rather than cap usage, the company changed how requests were routed across models and brought the bill down to 15 percent while holding steady at 600 billion tokens a month. It has now productized the dashboard it built to find the problem as the AI Spend Console, which scores prompts per day against shipped work and flags engineers whose output their peers keep sending back.

Bloomberg's Mark Gurman describes a screenless speaker roughly the size of a hockey puck, with a camera and sensors to read the room, a battery for portability, and moving parts intended to give it some personality. The reported price is $300 to $400, well above most premium smart speakers, for a launch in 2027. It is the first product out of OpenAI's $6.5 billion acquisition of io, the hardware studio founded by former Apple designer Jony Ive, and the pricing is the real question — a voice assistant is the cheapest category in consumer hardware, and this one costs about as much as a game console.