Anthropic Claude Opus 4.8: Features and Workflow Updates

Dynamic Workflows and API Improvements

Anthropic introduced dynamic workflows in Opus 4.8, allowing the model to spawn hundreds of parallel sub-agents to handle long-running, verifiable tasks such as large-scale code migrations. This system enables the model to write orchestration scripts that execute and verify sub-tasks before returning a final result. Additionally, the Messages API now supports system instructions directly within the message array. This change allows developers to update instructions mid-task without invalidating the prompt cache or routing updates through a user turn, significantly reducing costs for iterative development.

Control and Pricing Adjustments

Anthropic has replaced the previous adaptive thinking mode with manual effort control, allowing users to select from low, medium, high, extra high, and max settings. This change addresses community feedback regarding the lack of transparency and control over token budgets in prior versions. While standard pricing remains at $5 per million input tokens and $25 per million output tokens, the company has implemented a price reduction for Fast Mode, which is now three times cheaper while maintaining 2.5x speed improvements.

Benchmark Considerations

Performance evaluations for Opus 4.8 show competitive results on agentic coding benchmarks, though the author notes that benchmark scores are highly dependent on the specific harness used. For example, while Opus 4.8 scored 74% on the Agentic Terminal Coding Bench 2.1 using the standard harness, other models may report higher scores (e.g., 84.4%) when evaluated with different harnesses like the Codec CLI. The author emphasizes that developers should prioritize the specific harness configuration over raw benchmark percentages when evaluating model performance.