Claude’s latest model embraces transparency in its mistakes.

Anthropic is set to launch Claude Opus 4.8 on Thursday, emphasizing the model’s commitment to “honesty.” The company asserts that it trains all its models to be truthful, specifically aiming to avoid unsupported claims. However, it acknowledges a common issue with AI models: they can sometimes leap to conclusions, presenting their outputs confidently even when evidence is lacking.

Early testers have reported that Opus 4.8 is more adept at recognizing uncertainties in its outputs and is less prone to making unsupported assertions. In internal evaluations, it has been found to be about four times less likely than its predecessor to overlook flaws in the code it generates.

Beyond these honesty enhancements, Opus 4.8 introduces a feature that allows users to control the effort level Claude puts into tasks. Higher-effort responses will consume more tokens, giving users the flexibility to opt for lower-effort responses to manage their rate limits more effectively.

Additionally, Anthropic is unveiling a new feature called “dynamic workflows” in research preview, which is designed to enable Claude to tackle larger tasks. With dynamic workflows, Claude can plan its work and execute hundreds of parallel subagents in a single session, and with Opus 4.8, these agents can operate for extended periods. It also verifies its outputs before delivering results back to the user.