Anthropic to launch Claude Opus 4.8 with a focus on admitting mistakes

In short: Anthropic says it will release Claude Opus 4.8 on Thursday, and the company claims the AI will be better at saying when it is not sure or when it made a mistake.

What happened

Anthropic is launching a new version of its Claude AI model called Claude Opus 4.8. The company says it trains its models to be “honest,” meaning they should avoid stating things as facts when they do not have good support.

Anthropic says a common issue with AI tools is that they can sound confident even when they are wrong. The company claims early testers found Opus 4.8 is more likely to point out uncertainty in its own work and is less likely to make unsupported claims.

Anthropic also says its internal testing shows Opus 4.8 is about four times less likely than the previous version to let mistakes in code go by without mentioning them. Code is the step by step instructions programmers write for computers, like a recipe for software.

Why it matters

Many people use AI chat tools for writing, research, and help with work. When an AI confidently gives a wrong answer, it can be hard to notice, like a friend who guesses but never says they are guessing. If Opus 4.8 really flags uncertainty more often, it could make it easier for users to double-check important details.

Anthropic is also adding a way to choose how much “effort” Claude puts into a task. Higher effort uses more “tokens” (small chunks of text the system counts, like words on a meter), which can matter for people who have usage limits.

Finally, Anthropic is introducing “dynamic workflows” in a research preview. The company says Claude can plan a big job, split it into many smaller tasks running at the same time (like hiring a large team for one project), and then check its own results before responding.

Source: The Verge AI