Anthropic launches Claude Opus 4.8 with honesty as its killer feature
4x fewer hidden flaws and honest uncertainty in complex coding tasks.
Anthropic has released Claude Opus 4.8, a large language model that prioritizes honesty over raw speed. The company reports that Opus 4.8 is 4x less likely than its predecessor to allow flaws in generated code to pass unremarked, and it's more prone to express uncertainty about answers. Shopify engineer Tom Pritchard noted the model's improved judgment in Claude Code—asking the right questions, catching its own mistakes, and pushing back on unsound plans before making large changes.
Opus 4.8 also introduces dynamic workflows in research preview, enabling Claude to orchestrate hundreds of parallel subagents per session. These subagents verify outputs before reporting back, making the system suitable for massive tasks like codebase migrations across hundreds of thousands of lines. Additionally, effort settings—previously available in Claude Code—now extend to Claude.ai and Cowork, letting users trade depth for speed. Higher effort yields more thorough reasoning; lower effort yields faster responses. Pricing remains unchanged for regular Opus, while fast mode gets cheaper.
- Opus 4.8 is 4x less likely to pass code flaws unremarked compared to Opus 4.7.
- Dynamic workflows can run hundreds of parallel subagents that verify results before reporting back.
- Effort settings now available in Claude.ai and Cowork, allowing users to control reasoning depth vs. speed.
Why It Matters
Honest, self-correcting AI reduces debugging time and enables safer autonomous coding at scale.