Anthropic proposes AI governance framework with 10²⁵ FLOP compute threshold and government deployment veto authority

Anthropic has published a policy framework titled "Policy on the AI Exponential," calling for tiered regulatory oversight of frontier AI with specific numerical thresholds and explicit government authority to block dangerous model deployments. The document pairs an AI governance proposal with a companion economic policy framework.

What's new

The framework defines an Advanced AI Framework applying to models meeting any of:

• Trained using more than 10²⁵ floating-point operations (FLOPs)
• Developed by companies earning over $500 million in AI-related revenue, or
• Companies spending over $1 billion on AI R&D

Requirements for covered developers include publishing safety reports, engaging independent evaluators, maintaining security practices, and accepting government enforcement mechanisms. Four catastrophic risk categories are addressed: biological weapons development, cyber vulnerabilities, loss of control, and automated AI R&D acceleration.

A companion Economic Policy Framework addresses worker preparation and equitable distribution of AI's economic gains.

Context

Anthropics framework draws on its Responsible Scaling Policy, which ties internal development decisions to capability thresholds. This document proposes analogous external governance structures.

The timing is notable: Anthropic filed a draft S-1 with the SEC on June 1, 2026, bringing the company under increased public scrutiny. Publishing a detailed governance proposal within days of the S-1 filing positions Anthropic as a constructive actor in regulatory conversations at a moment when investor and policy attention is concentrated on the company.

The specific compute threshold—10²⁵ FLOPs—maps to training runs broadly equivalent to today's frontier models from OpenAI, Google, and Anthropic itself. Under this framework, Anthropic would be subject to its own proposed rules. So would OpenAI and Google DeepMind.

Why it matters

AI policy proposals often trade in generalities. Anthropic's framework is specific enough that regulators could begin drafting legislation around it without major definitional work. The compute threshold is a measurable, reportable quantity that cloud providers already log.

The proposal for government deployment veto authority is the sharpest public position Anthropic has taken on regulatory enforcement. Previous industry frameworks—including the voluntary White House commitments and the EU AI Act's risk tiers—stop short of explicitly granting governments power to block specific model releases.

For enterprise customers, the framework also signals Anthropic's expectation about the near-term regulatory environment: that numerical thresholds will become enforceable, and that companies not building internal compliance infrastructure now will face a harder transition when legislation passes.