OpenAI and Anthropic share findings from a joint safety evaluation

Source: OpenAI Date: 2025-08-27 URL: https://openai.com/index/openai-anthropic-safety-evaluation

Summary

Joint publication from OpenAI and Anthropic sharing findings from a collaborative safety evaluation exercise — both companies applied their respective evaluation frameworks to each other’s models (or to shared test sets) to compare safety outcomes and identify areas of divergence. This is the first known instance of direct safety collaboration between the two companies.

Implications

Safety/industry thread. OpenAI and Anthropic jointly publishing safety evaluation findings is a significant industry cooperation signal — two direct competitors collaborating on safety research. This likely reflects pressure from AI safety advocates and regulators for frontier labs to coordinate rather than compete on safety. The findings themselves would reveal where their safety frameworks agree (creating industry norms) and disagree (revealing different risk tolerances). This collaboration pattern may evolve into formal safety standards that pre-empt government-imposed ones.