Anthropic and OpenAI share results of joint AI safety tests

Anthropic and OpenAI have carried out a first-of-its-kind joint evaluation of their AI models, testing each other’s systems for potential safety and alignment risks. In blog posts released on Wednesday, Aug. 27, both companies said they examined public models using stress tests designed to detect problems such as sycophancy, whistleblowing, self-preservation, and capabilities that could undermine oversight. OpenAI said the collaboration highlights how AI labs can cooperate on safety challenges, while Anthropic described it as an effort to strengthen evaluation practices and establish best-in-class standards. Anthropic’s report noted that OpenAI’s o3 and o4-mini reasoning models performed as well or better than its own in overall alignment. However, GPT-4o and GPT-4.1 showed concerning misuse risks, and both companies’ models struggled with sycophantic behavior. The newer GPT-5, released after testing, was not included. OpenAI, meanwhile, found that Anthropic’s Claude 4 models generally respected instruction hierarchies and showed awareness of uncertainty, helping them avoid inaccurate statements. They did less well in jailbreaking tests that probe model safeguards, and their performance on scheming evaluations varied sharply depending on the test. For the purpose of these trials, both companies temporarily relaxed external safeguards that usually limit potentially harmful outputs in order to examine models under stress. Each said subsequent releases—OpenAI’s GPT-5 and Anthropic’s Opus 4.1—show measurable improvements compared with earlier versions. The effort underscores the growing focus on AI alignment, the challenge of ensuring that powerful systems act consistently with human values and interests. With policymakers debating regulation and the risks of fragmented state-level AI rules, industry players are stepping up efforts to demonstrate responsibility and transparency. By publishing their results and acknowledging weaknesses, Anthropic and OpenAI are signaling that cooperation may be necessary to address safety risks in increasingly capable AI systems, even as competition between labs intensifies.