OpenAI and Anthropic came close earlier this year to a legally binding agreement that would allow each company to stress-test the other’s commercially available artificial intelligence models, according to a report by The Information.
The proposed arrangement would represent a more formal version of an unusual experiment the two rivals conducted in 2025, when their researchers applied internal safety evaluations to one another’s publicly available models.
A Proposed Pact Between Fierce Rivals
Lawyers for the companies discussed an agreement under which each side would receive application programming interface access to the other’s commercial models. The access reportedly would not have extended to unreleased models. Tests would instead probe products already offered commercially for vulnerabilities, unexpected conduct and risks that a developer’s own evaluation process might overlook.
Data handling was also part of the proposed framework. Both companies would agree not to retain the other’s data generated or obtained during the testing process. Such a restriction would be especially sensitive in a relationship between competitors pursuing many of the same enterprise customers, researchers and computing resources.
The discussions are notable because OpenAI and Anthropic compete at the leading edge of commercial AI while presenting different approaches to product development and safety. Granting a rival structured testing access would introduce an uncommon form of peer scrutiny into an industry where the most revealing information about model design, evaluation and failure modes is usually kept inside individual laboratories.
The negotiations should not be treated as a completed partnership. The companies reportedly came close to an agreement earlier in 2026, but it remains unclear whether it was signed. Neither OpenAI nor Anthropic has issued an announcement confirming a final deal.
An Earlier Trial Established the Groundwork
The idea did not emerge from a blank page. In June and early July 2025, the two companies conducted what OpenAI described as a first-of-its-kind joint evaluation. Each laboratory ran selected internal safety and misalignment tests on the other company’s publicly released models, then published separate accounts of the findings in August.
OpenAI evaluated Claude Opus 4 and Claude Sonnet 4, while Anthropic tested GPT-4o, GPT-4.1, o3 and o4-mini. The work examined difficult areas including instruction hierarchy, jailbreaking, hallucination and simulated scheming. Some external safeguards were relaxed to prevent them from blocking the tests, although access still took place through public APIs rather than privileged access to internal systems.
Anthropic’s account of the exercise focused on tendencies including sycophancy, self-preservation, assistance with harmful requests and attempts to undermine safety evaluations. Anthropic said OpenAI’s o3 and o4-mini reasoning models performed as well as or better than its own models overall in the simulated tests, while GPT-4o and GPT-4.1 produced some concerning examples, particularly around misuse. Most models studied by both developers struggled to some degree with sycophancy.
Both companies warned against turning the results into a simple ranking. OpenAI said differences in access and familiarity made exact comparisons difficult, while Anthropic noted that API-level testing did not reproduce the additional instructions and safeguards used in consumer products such as ChatGPT and Claude. The exercises were deliberately adversarial and were designed to reveal possible behavior under pressure, not to estimate how often the same behavior would appear in normal use.
What a Formal Agreement Could Change
A binding arrangement would move the relationship beyond a limited research collaboration. Recurring access could allow each company to apply newly developed evaluations to the other’s commercial systems as capabilities change, potentially exposing weaknesses that are difficult for teams accustomed to their own models and testing assumptions to recognize.
Outside scrutiny, however, is only as valuable as the access and disclosure rules behind it. API testing can reveal how a model responds to carefully constructed prompts and simulated environments, but it does not automatically provide visibility into training data, internal reasoning records, system-level safeguards or how a company responds when researchers identify a serious problem. The reported proposal has not been published, so its testing scope, reporting requirements and procedures for resolving disputed findings remain unknown.
Commercial considerations would sit alongside the safety case. A reciprocal arrangement could demonstrate that frontier laboratories are willing to accept scrutiny from technically capable outsiders, but each side would also need to protect confidential information and prevent the testing process from becoming a channel for competitive intelligence. The proposed no-retention provision appears intended to address part of that concern.
Public reporting would ultimately determine how much the exercise contributes to wider accountability. The 2025 pilot was useful partly because both laboratories released methods, limitations and findings for researchers to examine. A private exchange may improve the companies’ own models, but published summaries, clear disclosure standards and a record of corrective action would make it easier for customers, regulators and independent evaluators to assess whether the process produces more than reassurance.
