Claude Opus 4.8: Anthropic Focuses on a More Transparent AI

Anthropic is set to launch its latest artificial intelligence model, Claude Opus 4.8, this Thursday, emphasizing a key feature: honesty. The company claims that its models are designed to avoid making unfounded assertions, a common issue in the AI field where models can sometimes draw hasty conclusions.

Initial feedback from testers indicates that Opus 4.8 is more likely to report its uncertainties and avoid unsupported claims. According to internal evaluations from Anthropic, this new model is about four times less likely than its predecessor to overlook errors in the code it generates without flagging them.

In addition to these improvements in honesty, Claude Opus 4.8 introduces a new feature that allows users to control the amount of effort the model dedicates to a task. Responses that require more effort will use more tokens, thus providing the option to choose less demanding responses to preserve rate limits.

Anthropic is also offering a new feature in research preview, called "dynamic workflows." This innovation allows Claude to handle even more complex tasks. With these workflows, Claude can plan work and execute hundreds of sub-agents in parallel during a single session. With Opus 4.8, these agents can operate for longer periods, and the model then checks its results before presenting them to the user.