Anthropic Releases Claude Opus 4.8 with Effort Controls and Dynamic Workflows
Anthropic has officially launched Claude Opus 4.8, building on the previous Opus 4.7 model with improved benchmarks, enhanced honesty, and a suite of new features aimed at complex agentic workflows.
The update introduces several novel features for end users and developers:
- Effort Control: Users on claude.ai and Cowork can now manually control how much effort Claude puts into a task.1 Higher effort settings prompt the model to "think more frequently and more deeply," while lower effort settings prioritize speed and conserve token rate limits.
- Dynamic Workflows: Available in research preview within Claude Code, this feature allows the model to spin up and orchestrate hundreds of parallel subagents in a single session to handle codebase-scale migrations and verify outputs2.
- Improved Honesty: Anthropic reports that Opus 4.8 is four times less likely than its predecessor to let bugs or flaws in generated code pass unremarked.
- Mid-Task System Message Updates: Developers can now inject system instructions mid-task within the Messages API array without breaking the prompt cache.
Importantly, Anthropic also teased the upcoming general release of its next-generation ultra-frontier model class, currently in testing under Project Glasswing.
-
An instance of Dynamic reasoning controls shift the cost of intelligence to adjustable inference-time compute. — Anthropic's latest model upgrade commoditizes model tiers by allowing users to actively dial computational overhead up or down on demand. ↩︎
-
An instance of Recursive self-improvement collapses the capacity of human developers to audit the resulting codebases. — Using hundreds of parallel agents to run massive codebase migrations requires humans to cede the code verification process entirely to the AI. ↩︎