Anthropic AI Admits When It Doesn't Know

Anthropic has officially released Claude Opus 4.8, its newest flagship AI model, aimed at fixing a critical AI flaw: overconfidence. The company states Opus 4.8 is substantially more truthful and less prone to fabricating information. Testers reported that interacting with Opus 4.8 felt akin to collaborating with another person, as it now poses clarifying questions and challenges nonsensical plans, a departure from previous models that often acted as 'yes-men'.

This advancement is significant as AI companies race towards autonomy. Opus 4.8 demonstrates notable improvements in legal reasoning, coding, and long-form analysis, with some users claiming it surpassed existing models in specific tasks. A new feature allows users to adjust the AI's 'thinking effort,' enabling it to spend more time on complex responses or prioritize speed and lower token usage. Anthropic has defaulted Opus 4.8 to high effort for optimal quality.

Anthropic also introduced Dynamic Workflows, enabling Claude to manage hundreds of parallel subagents for complex tasks, verifying and consolidating results. Claude Code can now manage massive codebase migrations. The company is also developing 'Claude Mythos Preview' for cybersecurity, with plans for even more intelligent models on the horizon, necessitating enhanced safety measures.