Google DeepMind unveiled Gemini 4 Argon, its new top model for coding, legal and finance work, and cyber defense. It can produce up to a million tokens in a single answer. Only invited testers can use it for now.

Anthropic's IPO prospectus casts AI as a bigger economic shift than the internet, and shows 2025 revenue of $4.6 billion, a $42 billion net loss and $518 billion in future compute commitments. The listing could value it above $2 trillion.

Anthropic launched Claude Sonnet 5.5, scoring 56 on the Artificial Analysis Index to rank #2 behind Opus 5.5. It matches leaders on some tasks but outputs far more text per task, raising cost to $7.60 per task despite unchanged pricing.

Tavus says Griffin is the first AI to pass a video Turing test, with 48% of testers mistaking it for human. It can listen, watch and talk at the same time, and is not yet public due to safety work.

NVIDIA, with over 100 partners, introduced the Open Agent Safety Platform, pairing OpenShell, open-source software that traces AI agents’ actions and enforces rules, with NVIDIA Sentry. The goal is to set clear, enforceable boundaries for agent behavior.

David Robinson, who wrote safety reports for OpenAI’s model launches, resigned and warned in The Atlantic that Silicon Valley’s sprint culture builds ever-bigger AI while ignoring risks, a problem deeper than new rules.

Microsoft AI released MAI-Transcribe-2-Streaming, a live speech to text model that ranks first on Artificial Analysis' streaming test with a 2.5% word error rate 0.13s after speech ends. It costs $9 per 1,000 minutes, pricier than ElevenLabs and Deepgram.

Ramp's latest AI Index shows business spending on AI APIs is falling, as frontier price cuts and cheaper, more efficient standard and lite models lower the cost of use. Spend peaked in July; open-source models remain under 5%.

Andrej Karpathy says we’ll spend more time interpreting language model answers and shares tips, including asking the model to explain in ASD-STE100, a controlled technical English standard originally developed for aerospace, to force clearer, simpler output.