Anthropic is set to release Claude Opus 4.8 on Thursday, boasting enhanced ‘honesty’—a term that now means admitting uncertainty and avoiding unsupported claims.

According to Anthropic, models are trained for honesty, but AI remains prone to jumping to conclusions. Early testers report that Opus 4.8 is more cautious about its work, flagging uncertainties more frequently than before.

The new model also offers users the option to specify how much effort Claude puts into a task—higher-effort responses use more tokens, giving users control over their rate limits. Additionally, dynamic workflows allow for more extensive tasks by running hundreds of parallel subagents in a session.

While this move towards honesty is a step forward, it raises questions about the reliability and trustworthiness of AI in various applications. Will we still believe its answers if it admits to not knowing something?