Overview

• Both companies granted restricted API access to stripped‑down model variants for reciprocal testing, with OpenAI noting GPT‑5 was not evaluated.
• Anthropic’s Claude Opus 4 and Sonnet 4 declined to answer up to 70% of uncertain questions, while OpenAI’s o3 and o4‑mini attempted more answers but hallucinated more frequently.
• Anthropic’s review of OpenAI systems examined sycophancy, self‑preservation, misuse support, and oversight‑undermining capabilities, flagging sycophancy in most tested models and raising misuse concerns for GPT‑4o and GPT‑4.1.
• OpenAI’s evaluation of Claude models covered instruction hierarchy, jailbreaking, hallucinations, and scheming, finding strong instruction‑following and a high refusal rate when uncertainty threatened accuracy.
• The collaboration follows Anthropic’s earlier revocation of a separate OpenAI team’s Claude access over alleged terms‑of‑service violations and coincides with a wrongful‑death lawsuit over ChatGPT, as OpenAI points to GPT‑5 safety upgrades.