What’s new
AI moves fast. Here’s what changed.
Model releases, open weights, new ideas and business news, in short. Every item links to the original and, where one fits, to the lesson that explains the idea behind it.
Last reviewed Oct 31, 2026 · Updated weekly
- Research & ideas
Trending: models that only use part of themselves
Mistral Large 4 is listed at about 1 trillion parameters with 52 billion active, and Meta’s new open model is reported at 816 billion with 42 billion active. Both quote a total and an “active” size, the mark of a mixture-of-experts design: only a few expert sub-networks run for each token, so a huge model costs far less to run.
Research & ideasGoogle gets a first AI response from a chip in orbit
Google Research says its Project Suncatcher prototype satellite sent back the first AI response from a data-center-class TPU in space, which the team calls its “Hello World” milestone. It is an early test of running AI compute off the ground.
- Policy & safety
OpenAI on how websites can signal AI training preferences
OpenAI published its approach to making it easier for website owners to say whether their content may be used to train AI. The question of who controls training data is one to watch if you build on or publish web content.
- Policy & safety
OpenAI publishes an updated Model Spec
OpenAI released a new version of its Model Spec, the public document describing how its models are meant to behave. The same week it also wrote about how it detects misalignment in frontier models.
Frontier modelsGoogle announces Gemini 4 Carbon
Google calls Gemini 4 Carbon its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. Long-horizon means staying on task over many steps, which is what agents depend on.
- Business
OpenAI and ServiceNow expand their partnership
OpenAI and ServiceNow announced a wider partnership covering enterprise workflows, one more sign of AI labs working through the software companies that businesses already use.
BusinessAnthropic files for an IPO
Anthropic filed an S-1 with the US SEC for a proposed initial public offering, planning to list on Nasdaq under the symbol ANT. The registration is not yet effective, and the number of shares and price range have not been set.
Open modelsDeepSeek posts V4.1-Flash-1017
DeepSeek added a dated update to its open V4.1-Flash model on Hugging Face. Dated checkpoints like this are how many labs ship small improvements to open models between big releases.
Open modelsMeta opens the weights of Muse Spark
Press reports say Meta released Muse Spark Open, an open-weight model of about 816 billion parameters with a 1 million token context window. It uses Meta’s own Muse Model License rather than a standard open-source one, so read the terms before building on it.
Frontier modelsMeta releases its first frontier models, Muse Wonder
Meta’s Superintelligence Lab released Muse Wonder through its API and Muse Wonder Full for Muse Code subscribers. By Meta’s own testing they roughly match the top models from OpenAI and Anthropic, though those results are not independent.
Frontier modelsAnthropic introduces Claude Haiku 5.5
Anthropic released Claude Haiku 5.5, the newest model in its small, fast Haiku line.
Open modelsMistral Large 4: previewed in October, weights now on Hugging Face
Mistral previewed Large 4, a natively multimodal mixture-of-experts model, on October 6 and promised the weights by the end of the month. Its Hugging Face page now lists them. Check the model card for the license before you use it.
Summaries are ours and kept short. Dates are when the source published. Benchmark and capability claims come from the companies themselves unless a source says otherwise, so treat them as claims, not proof. Spotted a mistake? Use “Report a broken link or error” in the footer.