Overview of Anthropic Claude Opus 4.8 Updates

Model Performance and Honesty

Claude Opus 4.8 demonstrates across-the-board improvements in benchmark performance compared to Opus 4.7, including gains in SweetBench Pro, multidisciplinary reasoning, and agentic financial analysis. While it remains slightly behind GPT 5.5 in terminal coding benchmarks, it shows a significant jump from 64 to 69 in agentic coding scores. A primary focus of this release is model honesty; Anthropic reports that Opus 4.8 is four times less likely than its predecessor to allow flaws in generated code to pass without flagging them. Furthermore, the model exhibits substantially lower rates of misaligned behavior, such as deception or cooperation with misuse, aligning its safety profile more closely with the Mythos model.

Workflow and API Enhancements

Anthropic introduced dynamic workflows, a feature designed to handle complex tasks by spawning tens to hundreds of parallel agents within a single session. Users can trigger this functionality by requesting a dynamic workflow in plain language or by enabling the 'ultra code' setting in Claude Code. Additionally, the Claude.ai interface and Co-work now include granular effort controls, allowing users to select response effort levels similar to those previously available in Claude Code. The Messages API has also been updated to accept system entries directly within the message array, enabling mid-task instruction updates. Notably, Opus 4.8 defaults to a 'high' effort level, though users retain access to two higher tiers, with increased rate limits in Claude Code to support the associated token consumption.