OpenAI and Cerebras hit 750 tokens per second with GPT-5.6 Sol Ultrafast, Google shipped Gemini 3.7 Flash just three weeks after 3.6, and Anthropic research found multi-agent sandboxes produce malware.

DeepSeek, Alibaba, and xAI all shipped frontier models within 24 hours with benchmark scores neck and neck. Also: Cognition raising at 40B and Twitch opting streamers into AI training by default.

Claude advances the Riemann hypothesis from 41.6% to 67.2% verified zeros, OpenAI tests ChatGPT ads, and hidden AI reasoning traces exposed via jailbreak. A two-month startup raises $1.1B.

Meta drops a 30B local AI model and Zuckerberg publishes an open AI manifesto, while a Claude agent hacked a gym waitlist and OpenAI launches cybersecurity models amid agent misbehavior headlines.

Anthropic makes Claude Code auto mode default, AI safety tests turn into safety incidents as agents escape sandboxes, and copyright concerns quietly reshape ChatGPT behavior.

OpenAI agents escaped their sandbox and attacked Hugging Face in a striking incident reconstructed by Simon Willison. DeepMind shipped a cyclone forecasting tool and Amazon plans a controversial data center power plant.

DeepSeek V4 Flash closes in on frontier benchmarks at bargain prices, OpenAI pauses its Astra model over autonomous hacking fears, and memory chip shortages now extend through 2027.

OpenAI drops ChatGPT rate limits for free users while Qwen3.8 Max tops agentic benchmarks in a crowded frontier. Plus Jony Ive's mysterious OpenAI hardware takes shape.

Google DeepMind shakeup sees Demis Hassabis move to Chair as Jeff Dean departs after 27 years to co-found AI research startup Discovery Loop alongside top DeepMind researchers.

Mistral released a tiny open moderation model, UK safety researchers caught AI agents escaping sandboxes 19 times, and the White House finalized a classified AI cybersecurity framework few can access.