Mistral announced the Mistral 3 family on 2 December: Mistral Large 3, a sparse mixture-of-experts with 41B active and 675B total parameters, trained from scratch on 3,000 NVIDIA H200 GPUs, plus three dense edge models at 3B, 8B and 14B. Every model ships under Apache 2.0 — Large 3 and the Ministral 3 series in base, instruct and reasoning variants. Mistral says Large 3 is its first MoE since Mixtral and debuts at #2 among non-reasoning open-weight models on LMArena, #6 across all open models.

Large 3 is multimodal for image understanding and multilingual, which Mistral claims is best-in-class for non-English, non-Chinese conversation; a reasoning variant is promised but not yet shipped. The Ministral tier is the cost half of the story: 3B to 14B models designed for offline and on-device use, fully fine-tunable, same permissive license. Coming one day after DeepSeek's MIT drop, that is a second permissive frontier family inside 48 hours.

Sam Altman sent a companywide memo on Monday 1 December, reported by The Information and viewed by the Wall Street Journal, telling staff that OpenAI was declaring a code red on ChatGPT and that we are at a critical time for ChatGPT. The surge concentrates on speed, reliability, personalization and answering a wider range of questions, with a daily call for those responsible and encouragement to move teams temporarily.

Work on advertising, health and shopping agents, and the personal assistant codenamed Pulse is pushed back as a result. The trigger is Gemini 3, released 18 November, which beat OpenAI's models on industry benchmarks; on users, Google reported 650 million monthly Gemini users in October against ChatGPT's more than 800 million weekly actives. The echo is exact: Google declared its own code red in 2022 when ChatGPT launched.

Anthropic announced on 2 December that it is acquiring Bun — the JavaScript and TypeScript runtime, package manager, bundler and test runner created by Jarred Sumner in 2021 — disclosing in the same post that Claude Code passed $1 billion in run-rate revenue six months after its May 2025 general availability. Terms were not disclosed. Anthropic confirmed Bun stays open source under the MIT license and that it will keep investing in it as a general-purpose toolchain.

Bun was already load-bearing for Anthropic: it powers Claude Code's native installer and has been the runtime under the CLI since mid-2025. An agentic coding session is a hot loop — resolve dependencies, execute, run tests, read output, iterate — so owning the runtime means owning cold-start latency and install behavior that, at a billion dollars of run rate, are unit economics rather than developer ergonomics.

DeepSeek published the official DeepSeek-V3.2 family on 1 December: 671B total parameters with 37B activated per token, a 128K context window, and weights on Hugging Face under the MIT license, alongside a technical report. The only architectural change from V3.1-Terminus is DeepSeek Sparse Attention, which turns attention cost from quadratic to near-linear in sequence length and roughly halves the price of long-context inference — the cut DeepSeek first attached to its September experimental build. A second checkpoint, V3.2-Speciale, spends far more post-training compute and reaches gold-medal-level results at IMO 2025 and IOI 2025.

The report frames V3.2 as reasoning-first and built for agents: it is the first model in the V3 line to fold thinking directly into tool use, and its reinforcement-learning data comes from a large-scale agentic task-synthesis pipeline rather than chat traces alone. DeepSeek's own positioning is a daily driver at GPT-5-level performance — worth reading literally, since GPT-5 shipped in August, meaning open weights are now chasing a three-month-old closed target rather than this month's one.