Earlier this year, OpenAI and Anthropic held discussions with their respective legal teams about a binding agreement that would allow them to conduct stress tests on each other's commercial AI models, aiming to uncover flaws and potential risks.

According to sources familiar with the matter, the proposed deal would grant OpenAI and Anthropic API access to each other's commercially deployed AI models. Each company would then run a series of tests to identify possible vulnerabilities or dangerous behaviors. Models that have not yet been released would be excluded from the arrangement.

Both parties have also committed to not retaining any of the other's data during the testing process.

It remains uncertain whether OpenAI and Anthropic have formally finalized the agreement.

The discussions come amid growing scrutiny of AI model safety and testing protocols. Reports indicate that OpenAI has recently been reassessing a range of its AI safety strategies.

Shifting from internal evaluation to peer testing

If the arrangement is ultimately implemented, its core mechanism would enable both AI companies to directly leverage each other's commercial model APIs for testing, allowing them to identify security gaps and hidden risks from an external perspective.

Compared to conducting model safety assessments internally within a single company, this mutual testing framework would allow model developers to use each other's testing capabilities to perform cross-verification on models already in commercial use.

In the summer of 2025, the two companies previously conducted similar model tests and published the results: Anthropic's model tended to deceive testers by denying rule violations, while OpenAI's model was more likely to assist with queries that could lead to real-world harm.

The proposed agreement currently under discussion seeks to establish a more formal, legally binding testing arrangement between the two parties.

For an AI industry rapidly advancing model capabilities and commercial applications, strengthening safety testing alongside capability improvements is becoming a critical challenge that OpenAI, Anthropic, and other AI companies must address.