Claude’s new model is more “honest” when he is wrong

Anthropic is launching the Claude Opus 4.8 on Thursday, and the company is touting the “authenticity” of the model.

According to To AnthropicIt trains “all of its models to be honest – for example, to avoid making claims they cannot support.” But she points out that “a general problem with AI models is that they sometimes jump to conclusions, confidently presenting their work as making progress despite little evidence.”

The AI ​​Lab claims that early testers found that Opus 4.8 was “more likely to signal doubts about its operation and less likely to make unsupported claims.” In the company’s assessments, Opus 4.8 is “about 4 times less likely than its predecessor to allow flaws in written code to go unnoticed.”

In addition to the honesty improvements, with Opus 4.8, users can guide the amount of effort Claude puts into a task. Higher effort responses will use a larger number of tokens, giving users the option of lower effort responses if they don’t want to exceed their rate limits as quickly.

Anthropic is also launching a feature called “Dynamic Workflow” in the search preview, which the company says will allow Claude to “take on bigger tasks.” Using a dynamic workflow, “Claude can plan the work and then run hundreds of parallel sub-agents in a single session (and with Opus 4.8, the agents can run longer). Then he verifies his output before reporting back to the user.”