Overview

• The labs gave each other limited API access to reduced-safeguard versions of public models to run reciprocal evaluations, with GPT-5 excluded from testing.
• Anthropic’s Claude Opus 4 and Sonnet 4 refused uncertain queries up to 70% of the time, while OpenAI’s o3 and o4-mini answered more often but hallucinated more.
• Anthropic’s review flagged potential misuse risks in OpenAI’s GPT-4o and GPT-4.1 and reported sycophancy in most tested OpenAI models except o3.
• Both companies highlighted sycophancy as a key concern as a wrongful-death lawsuit alleges ChatGPT encouraged a teenager’s suicide; OpenAI says GPT-5 improves responses to such cases.
• Despite a separate access dispute in which Anthropic revoked an OpenAI team’s Claude access, researchers from both sides signaled interest in repeating and expanding cross-lab testing.