AI

AI Labs Published Their Own Rulebooks This Week While Google Saved Its Strongest Model for Defenders First

By Crypt0 · Oct 29, 2026

The frontier AI labs spent this week writing down how their machines should behave, and that paperwork is becoming a product in its own right. OpenAI published a refreshed Model Spec on October 29, Google staged its strongest model for defenders first, and a fresh report on model distillation showed how quickly capabilities travel once they leave the lab. The thread tying the three stories together is proof. Labs are competing on published values, independent scores, staged access, and measurable capacity, and users get a clearer map of what each system is built to do.

OpenAI reorganized its Model Spec around five core values, faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Faithfulness to human intent sits at the foundation, asking models to understand what users, developers, and OpenAI intend and value, rather than following instructions in a purely literal way. The revised framework asks models to be transparent about what they know and what they leave open, and to help users make their own decisions and pursue their own goals, while staying subject to human oversight. OpenAI said current models are still growing into the updated Spec, with new training methods in exploration to help models internalize the principles. The move echoes Anthropic Constitutional AI approach, which uses a written constitution of principles to guide model behavior and make the values shaping AI systems more transparent.

Google DeepMind announced Gemini 4 Carbon on October 28, calling it its strongest model for cybersecurity and long horizon work. Carbon goes first to a subset of cyber defenders this week, then to more vetted organizations through the Fairwind Program, which counts more than 650 partners worldwide. Wider availability for developers, companies, and the public follows later, on a schedule Google has still to set, as part of a voluntary pre release testing process run by the United States government. Artificial Analysis ranks Carbon first on its Intelligence Index version 5.0 with a score of 59, ahead of Claude Opus 5.5 at 58, GPT 6 Astra at 57, and Gemini 4 Argon at 50. Google reports 88.2 percent on SEC bench Pro for finding vulnerabilities, 54.8 percent on ExploitGym, and 71.0 percent on CWE bench for patching, plus more than double the secure code success rate of Argon in internal tests, with unsupported claims roughly half the Argon level and high severity misalignment alerts down nearly 80 percent in live coding work.

The distillation story showed the other side of the proof race. Reuters, citing The Information and people with direct knowledge, reported October 29 that Chinese developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI GPT 6 Astra to train their latest models. Moonshot Kimi K3.1 Pro arrived last week, and DeepSeek upgraded V4.1 Flash 1017 launched October 17, with both models benefiting from distilling Astra, according to the report. OpenAI said in a September 30 blog post that it had linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI. Anthropic said earlier in October that new measures designed to prevent rivals from copying Claude capabilities had cut harvested data in distillation attacks by about 90 percent to 99 percent. Reuters reported responses from the three companies were still pending at publication, which keeps the report in the reported category and makes the technical debate about output controls even more central.

Capacity became the fourth headline, and it landed the same day as the Model Spec. OpenAI doubled usage limits for several paid tiers on October 29, with Work and Codex quotas doubling on the Pro 100 and Pro 500 plans, and the lower quota doubling on Pro 200 alongside a reversal of a planned cut after user pushback. The Plus plan keeps its current quota, new subscriptions already include the updated quotas, and existing customers receive them automatically. OpenAI framed the move around efficiency gains from its recently released GPT 6.1 Terra and Sol models, which match or beat their predecessors while using capacity more efficiently. The timing matters, with Anthropic preparing for its public listing and both companies competing hard for business users who want agents that can run longer, finish more work, and stay inside clear guardrails.

For builders and everyday users, the week offers a practical checklist. Read the published values before trusting an agent with real work, test a model on your own tasks before scaling it, and treat staged rollouts as a feature, since defenders get first access precisely because security work rewards careful deployment. The labs are learning in public, publishing scores, naming their values, and expanding capacity at the same time. That combination gives users more leverage than a single model launch ever could, because the next choice is clearer, compare the Spec, compare the independent score, compare the quota, and pick the system that earns its place in your workflow.

Sources

• Reuters, OpenAI updates Model Spec
• BetaNews, Gemini 4 Carbon to cyber defenders first
• Reuters, DeepSeek and Moonshot distillation report
• The Decoder, OpenAI doubles Pro usage limits

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