Google released a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode, positioned for science, research and engineering problems that “lack clear guardrails or a single correct solution.”

It rolled out to Google AI Ultra subscribers in the Gemini app, and for the first time became available through the Gemini API via an early-access program. Google reports 50.5% on CMT-Benchmark in advanced theoretical physics.

Anthropic closed a $30 billion Series G led by GIC and Coatue, valuing the company at $380 billion post-money — more than double its September valuation.

The company disclosed that Claude Code's run-rate revenue passed $2.5B, that weekly active Claude Code users doubled since January 1, and that enterprise use is now over half of Claude Code revenue. It cited an analysis estimating 4% of all public GitHub commits were authored by Claude Code. It also opened a research preview of a Model Hardware Standard, a shared spec for agents operating physical devices.

OpenAI's Ryan Lopopolo published “Harness engineering: leveraging Codex in an agent-first world,” describing five months of building an internal product with zero hand-written code: roughly a million lines, about 1,500 merged PRs from a team of three engineers driving Codex, delivered in an estimated one-tenth the time of hand-writing it.

A day later Can Bölük's “The Harness Problem” benchmarked 16 models across three edit-tool formats and found the format, not the model, drove the results: Grok Code Fast 1 went from 6.7% to 68.3% after switching off patch-style edits, MiniMax more than doubled, and Grok 4 Fast's output tokens fell 61% once it stopped burning tokens on retry loops. Patch failure rates on models not trained for Codex-style diffs reached 50.7%.

Z.ai released GLM-5 under the MIT license, scaled from GLM-4.5's 355B to 744B parameters with pretraining data up from 23T to 28.5T tokens, integrating DeepSeek Sparse Attention to cut deployment cost. It was trained on “slime,” an asynchronous RL infrastructure.

A day later MiniMax shipped M2.5, a 229B MoE with 10B active, posting 80.2% on SWE-Bench Verified and beating Claude Opus 4.6 on third-party harnesses. MiniMax reports it runs for about $1/hour at 100 tokens/sec and now writes 80% of newly committed code inside the company.

xAI co-founders Tony Wu and Jimmy Ba resigned within roughly 48 hours of each other, leaving Musk's lab with half of its original twelve. Reuters reported the exits are the latest in an exodus building over the past year; the Financial Times reported Ba's resignation followed tension inside the technical team over demands to improve model performance.

The departures came days after SpaceX said it would purchase xAI to create a $1.25 trillion company planning to go public later in 2026.

Mrinank Sharma, who led Anthropic's Safeguards Research Team, published his resignation letter on X, citing AI, bioweapons and “a whole series of interconnected crises,” and saying organizations “constantly face pressures to set aside what matters most.” He said he was moving back to the UK to study poetry.

The same day, OpenAI researcher Zoë Hitzig resigned as the company began testing ads in ChatGPT, and set out her reasons in a New York Times guest essay about manipulation risk. Two days later OpenAI dissolved the Mission Alignment team it created in September 2024, reassigning its seven members and giving lead Josh Achiam the new title of chief futurist — its second dedicated safety or mission unit stood down since 2024.