OpenAI and Anthropic Discuss Mutual Stress-Testing of Commercial AI Models

OpenAI and Anthropic are in discussions about a legally binding agreement that would give each other API access to their commercial AI models for stress-testing, according to The Information. The companies would run a series of tests to identify model vulnerabilities and potential dangers. It is not yet clear whether the agreement will be finalized.

People familiar with the matter said the two companies began discussing the arrangement with their lawyers earlier in 2026. Under the proposed agreement, testing would be limited to commercially released models, excluding unreleased ones. Both parties have also committed not to retain each other's data during the testing process.

The talks come amid growing attention to AI model safety and testing mechanisms. The Information reported that OpenAI has recently been reconsidering a range of AI safety policies. Unlike internal evaluations by a single company, this mutual testing mechanism would allow the two model developers to cross-check models already in commercial use using each other's testing capabilities.

In the summer of 2025, the two companies conducted similar model tests and published the results: Anthropic's model was more inclined 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 now seeks to formalize such testing arrangements in a more structured and legally binding manner.

For an AI industry rapidly advancing model capabilities and commercial deployment, strengthening safety testing alongside capability development has become an important issue for companies including OpenAI and Anthropic.