The New Claude Opus 4.8 Just Dropped — It Was Trained to Be More ‘Honest’ and Stop ‘Jumping to Conclusions’
The company says the latest version of its flagship AI model knows to admit when it doesn’t know something and stops making unsupported claims.
Anyone who uses generative AI chatbots knows they can sometimes be a little too confident with their answers. Anthropic, the maker of Claude, is aware of this and is trying to fix it.
The company released Claude Opus 4.8 on Thursday, the latest version of its flagship AI model. Anthropic says the upgrade was trained to be more “honest” — meaning it’s better at flagging when it doesn’t know what it’s talking about rather than confidently bluffing. In Anthropic’s own tests, Opus 4.8 is about four times better than its predecessor at catching its own coding mistakes.
Opus 4.8 also powers a new “dynamic workflows” feature in Claude Code that can run hundreds of parallel subagents on massive coding projects. Anthropic’s alignment team said the new model “reaches new highs” on supporting user autonomy and acting in users’ best interest. Pricing is unchanged from Opus 4.7: $5 per million input tokens and $25 per million output tokens.
Anyone who uses generative AI chatbots knows they can sometimes be a little too confident with their answers. Anthropic, the maker of Claude, is aware of this and is trying to fix it.
The company released Claude Opus 4.8 on Thursday, the latest version of its flagship AI model. Anthropic says the upgrade was trained to be more “honest” — meaning it’s better at flagging when it doesn’t know what it’s talking about rather than confidently bluffing. In Anthropic’s own tests, Opus 4.8 is about four times better than its predecessor at catching its own coding mistakes.
Opus 4.8 also powers a new “dynamic workflows” feature in Claude Code that can run hundreds of parallel subagents on massive coding projects. Anthropic’s alignment team said the new model “reaches new highs” on supporting user autonomy and acting in users’ best interest. Pricing is unchanged from Opus 4.7: $5 per million input tokens and $25 per million output tokens.
