Mistral Opens Preview of Large 4, Its Largest AI Model

Mistral, the French AI company, put Mistral Large 4 into public preview on October 6, 2026. Mistral says the model has about 1 trillion parameters, a rough measure of size, and calls it the strongest downloadable model outside China. For now it runs only through Mistral's paid service. The launch price is half the normal rate: about 68 cents for the model to read roughly 750,000 words. The files come out October 27, after cybersecurity experts finish testing them. Then companies can run Large 4 on their own computers. That matters to European firms that don't want their data going to American or Chinese providers. Mistral's announcement ↗

Anthropic Opens Its Strongest Cyber Models to More Security Teams

Anthropic expanded its Cyber Verification Program on October 6, 2026. Security teams that pass a check can now use Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5 for hacking-related work. There are now three tiers: defenders, red teams hired to break into systems with permission, and specialized groups. Nothing changes in the regular Claude app. That same day, JPMorgan CEO Jamie Dimon told Bloomberg that cyber risk "went up 10-fold after Mythos." Anthropic's announcement ↗

Claude Cowork Moves New Tasks to Anthropic's Servers

Claude Cowork, Anthropic's desktop feature that works through tasks on your files, now runs new tasks in Anthropic's cloud for Pro and Max subscribers. Anthropic's Help Center page, updated October 6, 2026, says files Claude opens during those tasks are processed on Anthropic's servers. Tasks created before October 6 keep running on your own computer. So if you picked Cowork because your files stayed on your machine, that no longer holds for anything new. Check which folders you give it access to. Anthropic Help Center ↗

OpenAI Publishes 722 Math Papers From an Unreleased Model

OpenAI released 722 math manuscripts on October 6, 2026. All of them came from an internal model the public can't use. The company gave the model about 4,000 problems, and each answer took roughly three hours of computing on average. OpenAI also posted versions of the proofs written in Lean, a language that lets a computer check a proof line by line, so mathematicians can test the work. Meta published six papers written with its Muse Spark model on October 4, so two companies have now used math research to show what their next models can do. OpenAI's post ↗

Cursor Lets Users Run Coding Agents From an iPhone

Cursor, the AI code editor, released a feature on October 6, 2026 that lets its iPhone app pair with Cursor on your computer. From the phone you can start a coding task or answer the agent's questions. The work stays on your own machine. If the phone loses signal, the agent keeps going. Cursor turned the feature on by default for every plan except Enterprise. Cursor changelog ↗

claude-mem Gives Coding Assistants a Memory Between Sessions

claude-mem gained 2,104 stars on GitHub this week. The free tool records what your coding assistant did in a session, uses AI to shrink that record into a summary, and feeds the relevant parts back the next time you start work. It works with Claude Code, Codex, Gemini and GitHub Copilot, among others. Read more ↗

open-dots Offers a Self-Hosted AI Agent Workspace

open-dots gained 901 stars on GitHub this week. You install the free workspace on your own server. It handles chat, links to other apps, and asks for your approval before an agent acts. Its author lists Claude Cowork and ChatGPT agent among the products it is meant to replace, and calls it an early prototype. Read more ↗