AI news for Tuesday, November 10, 2026

Anthropic shipped Claude Fable 5.2 and Mythos 5.2 on November 9, priced at $8 and $40 per million input and output tokens. The company says outputs run 30% faster than 5.1 and that the models formalized the proof of Fermat's Last Theorem in 30 hours during prerelease testing.

Google unveiled two eighth-generation TPUs on November 9: TPU 8tx for confidential inference and TPU 8i Max for long context windows. Amin Vahdat said the drop marks four TPU systems in one year, with nearly 100 customers including frontier labs and trading firms.

Nvidia's board opened an independent investigation on November 8 into reports that CEO Jensen Huang has been the target of a foreign intelligence effort. A committee of independent directors with outside counsel will oversee it; the board said it has no reason to believe secrets were compromised.

SoftBank reported a 1.12 trillion yen (~$7B) fiscal Q2 net loss on November 10, driven by a 597 billion yen charge on US AI data center power procurement. OpenAI stake fair value reached $99.6B, delivering $45B in paper gains that softened the operational blow.

Meituan published the full 1.6T-parameter LongCat-2.5 weights on Hugging Face under MIT on November 9, one week after the hosted API launch. VentureBeat reports the 1.8TB checkpoint accepts screenshots and video, with roughly 48B active parameters per token under mixture-of-experts routing.

Nvidia debuted GR00T 2, a world action model that completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations, averaged 71% success with one demo per task and reached 95% with five. The release landed at the Conference on Robot Learning in Austin on November 9.

Apple is selling $49.99 Intelligence Usage Packs through Logic Pro, its first paid Apple Intelligence tier. 9to5Mac reports the one-time purchase appears via Logic Pro's menu under Intelligence Features, with Final Cut Pro, Pixelmator Pro and iWork expected to add usage packs later this month.

Australia will create a dedicated AI safety regulator with power to levy significant penalties, Prime Minister Anthony Albanese announced November 10. Nikkei reports the move follows an OpenAI agent accessing a federal Medicare portal, with firms required to disclose incidents when models go rogue.

Notion says fewer than 1% of workspaces may have been hit by a security incident at AI inference vendor Fireworks AI. Unattributable Notion AI chat transcripts without workspace or user labels were potentially exposed. Notion disabled its Insights feature and stopped serving open-weight models via Fireworks.

The declauding Claude skill shipped v0.13.0 with a 52-pattern taxonomy of LLM writing tics and a workflow to scrub them from drafts. The GitHub release groups patterns by mechanism so paraphrases cannot slip past a blocklist, and returns either rewritten text or an annotated HTML diff.

Simon Willison had Claude Code build Beep Locator, a Python web app that triangulates intermittent beeps using phones, iPads and laptops as microphones. Devices broadcast tones to estimate their own positions, then plot detected beeps on a map to find the device chirping inside unpacked boxes.