Anthropic, Google, and OpenAI Launch Frontier Models in Dual 'Public-Vetted' Releases; Safety Guardrails and Access Controls as Key Differentiators

In early September 2026, Anthropic, Google, and OpenAI introduced a strategy of releasing their top-tier AI models in two versions: a 'public version' and a 'vetted version.' Anthropic defined this as 'the same model but with different levels of safety guardrails.'

Dual-Release Structure and Pricing Policy

The public version is available to general developers via API but may have restrictions on certain features or usage patterns. In contrast, the vetted version has no public price list and is accessible only to government agencies, critical infrastructure operators, or organizations in specific fields (e.g., cyber defense, life sciences) that have passed strict review.

• Anthropic: Released Claude Fable 5.1 (public) and Claude Mythos 5.1 (vetted). The price for Fable 5.1 is $10 input/$50 output per million tokens, exactly double that of the previous model, Opus 5. Mythos 5.1 is available only to a small number of vetted organizations in the US, with the approval process linked to the Organization ID.
• Google: Distinguished between Gemini 3.8 Flash (public) and Flash Cyber (vetted). Flash maintains a promotional price of $0.75/$3.75 until the end of the year. Flash Cyber is provided only to government and cyber authorities, critical infrastructure operators, etc., through the Fairwind Program, with no pathway for individuals to sign up.
• OpenAI: Released GPT-6 Astra (public). The price is $10/$50 per million tokens, the same as Fable. Performance of the vetted version is first provided to alpha testers. OpenAI requires a government-issued ID and two hardware security keys for individual users.

Developer Impact and Precautions

Public and vetted versions may share the same weights, but differ in safety guardrails and monitoring levels. Notably, OpenAI's public API applies a monitoring system that detects and automatically interrupts 'long-running agent tasks.' Therefore, developers should be aware that benchmark scores are measured on the vetted version and consider these limitations when using the public API.