This summer, OpenAI and Anthropic collaborated on the first-ever cross-company safety evaluation, testing each other’s publicly released models, including Claude Opus 4, Claude Sonnet 4 (Anthropic) and GPT-4o, GPT-4.1, o3, and o4-mini (OpenAI), by running internal misalignment and misuse assessments.
The tests explored key behaviors such as sycophancy, misuse potential, hallucinations, and resistance to instruction. Both parties discovered safety gaps: although reasoning models (like o3 and Claude’s reasoning-capable models) generally exhibited stronger alignment, general-purpose models like GPT-4.1 proved more vulnerable to misuse.
These early findings emphasize the need for continued collaboration and rigorous testing in AI safety.
