OpenAI and Anthropic negotiated a legally binding deal earlier this year to stress-test each other's commercial AI models, The Information reported Monday, with the New York Post, Mint, and SBS English retelling the same account. If those talks became a live contract, the two frontier labs would have put peer red-teaming into enforceable terms instead of leaving safety probes inside each company alone. Public rewrites still say it is unclear whether anything was finalized.
The reported pact is not a merger, a joint product, or a shared training run. Under the proposed terms, each side would get API access to the other's commercially available models to hunt safety risks and unexpected behavior, and would agree not to retain testing data. The Hacktron staff-account disclosure and British Columbia's Tumbler Ridge failure-to-warn suit are separate matters; this account is only the mutual stress-test near-deal.
Lawyers for both labs were drawing up the terms before high-profile cybersecurity incidents involving OpenAI models, including the Hugging Face episode cited in the Post's rewrite, according to a person with direct knowledge quoted via The Information. The companies previously ran a lighter informal joint safety evaluation in August 2025 with special API access, which Anthropic called a first-of-its-kind cross-lab test.
Representatives for OpenAI and Anthropic did not return comment to the Post; SBS said both declined to comment. The leak lands amid a wider public argument over industry slowdowns, third-party embedded evaluators, and peer review among labs, including comments from Elon Musk at the All-In Summit that top labs should test each other's models.
What remains open is whether any signed agreement exists, how far draft terms went, and whether antitrust concerns Amodei has raised about industry cooperation would block a formal bilateral pact. Until either company confirms, the story stays reported talks and a near-deal, not a closed contract.