AI

Control Became the Product as Open Weights, Rulebooks, and Opt Outs Landed Together

By Crypt0 · Oct 30, 2026

This week gave AI watchers a clear through line, and it runs through control. Mistral handed over downloadable weights for a trillion parameter flagship, OpenAI published a rewritten rulebook for how its models should behave, and the four largest Western labs agreed on a single signal publishers can use to shape AI training choices. Each move answers the same question from a different seat in the room, who gets to decide how powerful models are built, run, and used. The answers point toward builders and publishers holding more of the levers themselves.

Mistral opened the weights on October 29 for Mistral Large 4, nicknamed Le Chonk, a natively multimodal mixture of experts model with about 1 trillion parameters in total and about 49 billion active on each token, plus a 1 million token context window. The release followed three more weeks of reinforcement learning after the October 6 public preview, and Mistral reports gains across coding and defensive security work. Terminal Bench 4.0 rose to 41 percent, up 13 points from the preview, with output tokens down 30 percent. The Artificial Analysis Coding Agent Index came in at 48 percent, effectively tied with GLM 5.3 while using 47 percent fewer tokens per task. Cognition FrontierCode 2.0 reached 19.7 percent, the top open weight score Mistral reports. On security, Mistral reports 51 percent on the Artificial Analysis Cyber Index and a 96.6 percent refusal rate on malicious cyber requests across JailbreakBench, StrongREJECT, and AgentHarm. The weights ship under the Mistral Ecosystem Open Weight License, a modified MIT license that allows commercial use and fine tuning, while running the model as a paid service for others calls for a separate commercial license. Self hosting takes serious hardware, with an eight graphics processor server recipe for the main checkpoint and a compressed variant built for Blackwell graphics processors. Independent scoring from Artificial Analysis puts Large 4 at 26 points on its Intelligence Index version 5, the highest score among Western open weight models, with leading Chinese open weight models scoring higher.

OpenAI published its updated Model Spec on October 29, the public document describing how its models should behave, and reorganized it around five values, faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The revision adds guidance for a world where models take actions on their own, covering proportionate use of context, memory, and personalization, respect for privacy and user control, support for human agency and oversight, social norms, and clear communication about side effects. Much of the detailed style guidance retired, with style now adapting to what each person prefers, and the model is called GPT throughout the document. OpenAI says it is exploring training methods that teach models why the principles matter, alongside what the principles say, with results to share later. The Spec leaves today's ChatGPT capabilities exactly as they are, and that is precisely the point, it states plainly what OpenAI wants its models to do tomorrow, in public, where builders and users can read it and hold the company to it.

October 30 brought the coordination story into full view, as OpenAI, Google, Microsoft, and Anthropic said they will honor a single robots.txt policy token for AI training, letting a website express one training preference across every participating provider at once. A site owner writes the token as the user agent line, followed by a disallow line for the whole site, with allow lines carving out exceptions such as a press kit. Anthropic says it has already started implementing support, OpenAI aims to implement the signal in the coming months, Microsoft targets February 1, 2027, and Google targets 2027. The token sits inside an individual Internet Draft titled Machine Readable Training Preference Signal, authored by Anthropic engineer Kevin Kelley and dated October 30, 2026, with an expiry of May 3, 2027. An individual Internet Draft carries the status of a proposal on the path toward a formal standard, and the AIPREF working group chairs come from Cloudflare and Cisco. The token shapes training choices, while fetching stays separate, so crawlers may still pull a page for search and other uses, and material already collected stays in existing datasets. OpenAI frames the mechanism as a coordination tool, which keeps expectations honest about what one line in robots.txt can accomplish on its own.

Read together, the three moves sketch a maturing market where openness, published principles, and shared signals reinforce each other. A company can run a frontier class model on its own servers, a publisher can state one training preference in one place, and a user can read the values a model is meant to follow. The work ahead sits in implementation, independent testing of the benchmark claims, and the slow spread of the new token across the web. For readers, the practical payoff is tangible, more choice about where models run, clearer language about how they should behave, and a simpler way for publishers to speak for their own work. The frontier keeps getting more capable, and this week it also got more legible.

Sources

• Mistral Large 4 open weights release
• Mistral Large 4 weights and license details
• Four labs back one AI training opt out token
• How the shared training signal works

New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.