Claude’s new model is as ‘honest’ as it is confusing

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

According to at Anthropicit teaches “all (its) examples to be honest — i.e., to refrain from making statements that they cannot support.” But it says “the biggest problem with AI models is that they sometimes jump to conclusions, confidently suggesting that their work is progressing despite little evidence.”

The AI ​​Lab says that early testers have found that Opus 4.8 “tends to show uncertainty in its performance and that it will not say anything that does not agree with it.” In the company’s assessment, Opus 4.8 “is about 4x less likely than it was originally designed to allow for written errors to occur without reference.”

In addition to the fair changes, with Opus 4.8, users can control the amount of Claude’s input. Advanced solutions will use more symbols, giving users the opportunity to reduce the number of solutions if they don’t want to burn their limit too quickly.

Anthropic is also introducing a feature called “dynamic workflows” for viewing research, which the company says will allow Claude to “do more complex tasks.” With powerful workflows, Claude can prepare the work and then run hundreds of parallel operations in one session (and with Opus 4.8, agents can run longer).