Report: OpenAI and Anthropic pursued an agreement to mutually verify AI models

According to a report from the Korean broadcaster SBS, OpenAI and Anthropic pursued an agreement to mutually verify each other's AI models. The headline fact is narrow but pointed: two of the most advanced model developers looked at ways to put a rival's systems through their own verification process.

That is notable because frontier model evaluation is still largely self-assessment. Developers test their own systems or commission evaluations they pay for, which makes independent judgement of safety and robustness claims unusually hard, and leaves model cards and system cards resting on the company's own word.

Cross-checking by a competitor changes the calculus. OpenAI and Anthropic compete directly in assistants and enterprise AI, so letting a rival into the evaluation loop would buy more independence than self-reporting, while raising hard questions about trade secrets, model exposure and competitive risk.

The move fits a wider trust deficit. Model capabilities are advancing faster than the evaluation tooling, audit standards and third-party institutions meant to keep pace with them, and regulators are increasingly asking for disclosure that companies cannot easily substantiate.

Peer verification between rivals can also be stickier than a unilateral pledge. A promise can be quietly revised or withdrawn; a standing mutual-verification process makes withdrawal a signal in itself.

Three things to watch: whether the arrangement is actually implemented, which models and which metrics it covers, and whether third-party standards bodies or regulators are brought in. The underlying question is uncomfortable for the whole field: when the developer is also the primary source of safety claims, who verifies the verifier?

Even if the arrangement never takes shape, the fact that it was reported says something in itself: frontier labs have started to treat each other's models as objects that need outside verification, not merely as rivals to out-compete.