Mistral opens Large 4's weights, Gemini Computer Use goes GA, and one robots.txt line to opt out of AI training

Mistral Large 4 is a 1-trillion-parameter mixture-of-experts model (only some experts fire for each token) that handles both images and text. Going by the repo name, about 52B parameters are active per token. After the Oct 6 preview, Mistral ran three more weeks of reinforcement learning and has now released the improved weights in FP8. You can pick Low, Medium or High (the default) reasoning, and the repo is tagged for vLLM, SGLang and llama.cpp.

Mistral says it leads open-weight models on Cognition's private FrontierCode 2.0 at 19.7%Vendor claim and on Terminal-Bench 4.0 at 41%Vendor claim. On the Artificial Analysis Coding Agent Index it reports 48%, about level with GLM-5.3 while using 47% fewer tokens per taskVendor claim. It also says the model takes on vulnerability finding and patching without needless refusals, yet refuses 96.6% of malicious cyber requests on averageVendor claim. For teams that can't send code off-premises, that's one more serious option for a self-hosted coding or security agent.

Computer Use lets the model look at a screenshot and return the next action, such as a click, scroll or keystroke, as a function call. Your own code (Playwright, for example) performs the action and sends back a fresh screenshot, and the loop repeats. It's now generally available in both the Interactions API and the GenerateContent API. Supported models include gemini-4-argon, gemini-3.8-flash (the one the docs recommend) and gemini-3.5-flash-lite.

pip install google-genai playwright
playwright install chromium

With the preview label gone, it's easier to put browser automation into production: end-to-end tests for web apps, repetitive form filling, research across sites. It works in browser, mobile and desktop environments. Each action comes with an intent explaining why the model chose it, plus a safety_decision that can ask for user confirmation on risky steps. You can also opt in to prompt-injection detection that scans screenshots for hidden instructions.

Put the policy token __usage_ai-training in a robots.txt User-Agent line, where a crawler name would normally go, and the rules beneath it state whether those paths may be used for AI training. There's no new syntax; normal robots.txt path matching and longest-match precedence apply.

User-Agent: __usage_ai-training
Disallow: /
Allow: /press-kit/

Until now, site owners had to block crawlers one company at a time: GPTBot, Google-Extended, ClaudeBot and so on. As more companies honor this token, that collapses into a single group. Google says an opt-out through the token will count as disallowing Google-Extended, and existing per-company rules keep working. If you run a content or docs site, it's a good moment to revisit your robots.txt policy.

gemini-api is a stateless, zero-dependency command-line client. From inside a coding agent such as Claude Code, Codex or Cursor, or from a CI job, you can generate images, analyze files and run Gemini managed agents as plain shell commands. It's a different tool from @google/gemini-cli, which is an interactive coding agent.

curl -fsSL https://raw.githubusercontent.com/google-gemini/gemini-api-cli/main/scripts/install.sh | bash
export GEMINI_API_KEY="YOUR_API_KEY"
gemini-api analyze -i ./report.pdf "Summarize the three key points."

When agents call REST directly, base64 media floods the context window and interactive prompts stall subprocesses. This CLI writes output files to disk and prints only the path, and --dry-run, -o json and --agent-mode give machine-readable output and errors. It's a handy way to give an agent Gemini's image and research features without adding an SDK to your project.