The US says OpenAI and Anthropic agreed to let the US AI Safety Institute have early access to major new AI models to evaluate their capabilities and risks
The US government announced agreements with leading artificial intelligence startups OpenAI and Anthropic to help test and evaluate their upcoming technologies for safety.
Context & Ripple Effects
The agreements turn voluntary safety commitments into a concrete access arrangement: the US AI Safety Institute can examine major models before public release rather than relying only on developers’ internal assessments.
They sit at the start of a broader external-evaluation arc. OpenAI and Anthropic later reported joint safety testing of each other’s models, and subsequent coverage says Google, Microsoft, and xAI joined the US pre-release-access framework for government model evaluation.
First-order effects
- OpenAI and Anthropic must provide the US AI Safety Institute early access to major new models, giving the institute a pre-release channel to assess capabilities and risks.
- The institute gains a defined role in evaluating frontier systems before they reach the public, while the two labs add an external government reviewer to their release process.
Second-order effects
- The arrangement establishes external pre-release review as a practical benchmark for other major model developers; the later participation of Google, Microsoft, and xAI indicates that the channel can broaden beyond its initial two labs.
- Model developers and government evaluators must develop working processes for controlled access, testing, and communicating findings without making pre-release models broadly available.
Third-order effects
- If such arrangements become routine, access to unreleased frontier models could become a durable layer of AI governance, with leading labs differentiated partly by their ability to support state evaluation.
- The pattern may shift safety oversight from developer-only claims toward recurring review relationships between governments and a small set of frontier-model providers, though the influence of those reviews on release decisions remains unclear.
The trend: Frontier AI governance is moving toward institutionalized pre-release model access for state-backed evaluators.