This Week in AI, Nov 2–8, 2026

AI labs announced seven models and five other events in the week from Monday, Nov 2, to Sunday, Nov 8, 2026, and OpenRouter listed one stealth model, by Frontier Log's timeline of their official posts. All 13 came between Monday and Thursday, and seven of them came on Wednesday, Nov 4. The timeline records no event from Friday to Sunday.

Seven of the eight models came on Nov 4 and 5

OpenRouter listed Crouton Alpha on Nov 3, a stealth model for coding, agentic and multimodal work with a 1,048,576-token context window. OpenRouter says an anonymous third-party provider develops and operates it.

Four labs released models on Nov 4. Tencent launched Hy4, a new model that adds vision and computer use for the first time and costs the same through the API as the Hy4 preview. Tencent said it plans to open-source the weights. Ricursive Intelligence released Ripple-65B Preview, a 65B dense model trained from scratch with up to 1M tokens of context, with its weights on Hugging Face. Cognition introduced SWE-2 Flash, a smaller coding model built on Inkling-Small. OpenAI launched Images 2.6 in ChatGPT, Codex and the API.

Three more came on Nov 5. OpenAI introduced GPT-6 Cyber, its cybersecurity model, at first for a select group of approved Daybreak Red customers. Thinking Machines released Inkling 2 Small, a mixture-of-experts model with 594B total and 17B active parameters, through Tinker with closed weights at first. It said it intends to release the full weights under an Apache 2.0 license the following week, after external safety testing and red-teaming. Google launched Gemini Omni 1.2 Flash, which makes and edits videos up to 15 seconds long. It costs 30% less than Gemini Omni 1.1.

The other five announcements

Anthropic released FlashProve on Nov 2, an open-source research prototype for provable inference. It said a single H100 GPU generates proofs for Meta's Llama 4 Maverick, a 400B-parameter model at 4-bit, at over 100 tokens per second. That takes about 70 to 95 times the GPU time of processing the same prompts with vLLM.

On Nov 4, OpenAI posted three times. It reported that an autonomous pipeline found, validated and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel. Its alignment team reported on GPT-6 Astra in OpenAI's internal deployment. Astra completed about 10% of hidden rubric items on tasks framed as safety research. On tasks framed as capabilities research, it completed about 16%. OpenAI also said it will acquire Pierre Computer Company, subject to regulatory approvals.

On Nov 5, Anthropic began rolling out Personal Finance in Claude in beta on the Pro, Max and Ultra plans in the US. Its access to linked bank and card accounts is read-only.