The State of AI

Where things stand, November 3, 2026

The frontier now moves in weeks. In October alone, OpenAI shipped GPT-6.1 Terra, Google moved the Gemini 4 line from announcement to general availability and then announced a restricted successor, Anthropic filled out the Claude 5.5 family, and Meta launched Muse Wonder. Prices at the capable end keep falling: work that needed a flagship in August is mid-tier work now.

Open weights are closing the distance in the open. Mistral put the Large 4 weights up for download, DeepSeek kept shipping dated builds, and small open models like Liquid AI's LFM3 family now cover text, vision, and audio at sizes a phone can carry. The argument between open and closed is no longer about whether open models are good. It is about how many months behind they run.

The deepest change is what the models are doing. In October, two labs announced machine-produced proofs of a conjecture that had stood since 1999, alongside proofs in harmonic analysis and quantum field theory, formalized in Lean so other researchers can check them line by line. Add world models that generate sound and speech with the scene, and voice models that talk speech to speech, and the picture is a field moving from answering questions to producing work that other experts must now evaluate.

The first days of November added two harder edges to that picture. Meituan released a frontier-scale model trained entirely on Chinese chips, proof that the largest builds no longer depend on Nvidia hardware. Epoch AI showed a downloadable open model turning published security fixes into working exploits at about a thousand dollars a release, and China said it will speed up a comprehensive AI law covering a model's whole life cycle. Capability, misuse, and regulation are now moving on the same clock.

The legal and safety edges sharpened at the same time. An adviser to Europe's highest court drew the copyright line between training, which stays permitted, and answers built from retrieved text, which can infringe. Anthropic, amending its IPO filing, said its coming models may imminently meet the threshold it set for AI-driven acceleration of AI research itself, while stating plainly that no current model has crossed it.

The list

• Nov 3, 2026An EU court adviser says chatbot answers can infringe publishers' rights, while training stays permitted. Advocate General Maciej Szpunar of the Court of Justice of the European Union, in an opinion delivered November 3 in Like Company v Google Ireland (Case C-250/25), said that when a chatbot retrieves a press publication, reuses it to generate an answer, and provides that answer to the user, the publisher's right may be infringed in certain circumstances: notably where the access circumvented a paywall or other technical barrier, or where the answer reuses the publication's form of expression in a recognisable manner. Copyright protects that expression, not the mere facts a publication reports. Using protected material to train models, he said, is freely permitted under the EU's text-and-data-mining exception, and there was no evidence the Hungarian publisher's articles had been used to train Gemini. The opinion is advisory and not binding; the court's judges rule in the coming months, and they follow the majority of such opinions. Coverage at Reuters
• Nov 3, 2026China says it will speed up a comprehensive AI law covering the full life cycle. A senior official at the National Development and Reform Commission said the planned rules would cover data, computing power, and algorithms, set safety standards from development through deployment, and require disclosure of safety risks and reporting of incidents and vulnerabilities, with regulatory sandboxes and transition periods. The country's frontier developers, DeepSeek, Moonshot, and Alibaba among them, are in scope. This is an announcement of legislation to come, not a law in force. Coverage at Bloomberg Law
• Nov 2, 2026Anthropic's amended IPO filing says it may imminently meet its own self-improvement threshold. In an amended S-1/A filed November 2, Anthropic added risk-factor language saying its internal estimates for future models, not current ones, indicate it may imminently meet its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI," the threshold it associates with recursive self-improvement. The filing says the company has in the past used AI to develop models whose capability leaps it did not identify at the time, that it has not designated any current model as crossing the threshold, and that it is considering updates to the policy to keep humans in control of frontier model development. A risk factor is a disclosure to investors, not a claim that the line has been crossed. Coverage at TechCrunch
• Nov 2, 2026Meituan releases LongCat-2.5, trained on Chinese chips. The model is a mixture of experts with 1.6 trillion total parameters and about 48 billion active per token by Meituan's figures, a context window of one million tokens, and native multimodality, aimed at long coding and agent work. Meituan says it was trained entirely on domestically produced accelerators, its second model in a row built that way. It is live through Meituan's chat service and API at a discounted $0.30 in and $1.20 out per million tokens; the weights are not open yet, with a GitHub and Hugging Face release planned, and its benchmark claims are the company's own, with no independent evaluation published yet. Coverage at Tech Times
• Oct 30, 2026Epoch AI shows an open-weight model turning Firefox security fixes into working exploits. In Epoch's experiment, Z.ai's open-weight GLM 5.3 read security fixes that were public but had not yet reached users, and built working exploits for six Firefox releases at a marginal cost of about $1,000 per release, with no provider refusing the work. The fastest took about 12 hours. The attacks read files after a user opened a web page, and relied on a separate unpublished flaw the model found, which Mozilla fixed in Firefox 159.0 on October 27. Epoch has not confirmed the attack still works without that flaw, and a comparable attempt against Chrome failed. Coverage at Open Data Science
• Oct 29, 2026Mistral releases the Mistral Large 4 weights. The French lab's flagship goes open: after a public preview that opened October 6, the weights themselves are now downloadable. The frontier is no longer only rented by the token. The week in AI, Oct 24 to 30
• Oct 29, 2026OpenAI updates GPT-6 Luna toward shorter, less agreeable answers. A revised Luna ships in ChatGPT with a system-card addendum. Tuning a model to push back instead of flatter is now a release feature, which says something about what users complained about. The week in AI, Oct 24 to 30
• Oct 28, 2026Google announces Gemini 4 Carbon, cyber defenders first. The newest Gemini is not generally available: its first seats go to security teams. Capability releases are starting to be rationed by use case, not just by price tier. The week in AI, Oct 24 to 30
• Oct 28, 2026Anthropic publishes its own refutation of the Exponential Time Hypothesis. A second frontier lab reaches the same complexity-theory result OpenAI announced two days earlier. Two independent machine proofs of one long-standing conjecture is a story about verification as much as discovery. AI News Briefing, Oct 28
• Oct 27, 2026Google DeepMind previews Genie 4 to trusted testers. The world model builds an interactive world in real time as the user steers it, and now generates music, ambient sound, and character speech along with the scene. World models are becoming multimedia engines, not just video predictors. AI News Briefing, Oct 28

More developments

How this list is kept

Entries are added when a development is real enough to source: a release, a paper, a filing, a result. Rumors stay off the list until they land. Each entry carries the date of the development, not the date it was added here, and links the coverage or tracker it was drawn from. Where the keeper of this page has a stake, the entry says so: the Muse Wonder entry above is the standing example.