Claude Opus 4.8 Learns to Say “I Don’t Know” as Anthropic Prioritizes Honesty

Anthropic’s new Claude Opus 4.8 places a major emphasis on honesty, with the model showing near‑perfect performance in benchmarks that test whether an AI can admit uncertainty or say “I don’t know.” While the upgrade is modest in raw capability compared to Opus 4.7, it significantly improves transparency and reduces overconfident answers, a persistent issue in LLMs. Anthropic notes that Opus 4.8 even showed signs of “evaluation awareness,” reasoning about how its answers might be graded during testing. Mythos Preview remains more powerful overall, but Opus 4.8 is now the most forthright model in general availability.

Read the full story on PCWorld →