Claude Opus 4.8: Improved Honesty, Effort Control, and Dynamic Workflows

Anthropic has released Claude Opus 4.8, a new version that improves on Opus 4.7 across benchmarks while introducing features like effort control and dynamic workflows. The update addresses a common tension in AI models: balancing capability with honesty and reliability. Early testers report that Opus 4.8 is more likely to flag uncertainties and less likely to make unsupported claims—a roughly fourfold reduction in allowing flaws to pass unremarked compared to its predecessor. This focus on honesty makes the model a more trustworthy collaborator for complex, autonomous tasks.

Technically, Opus 4.8 defaults to high effort for optimal quality but lets users dial effort up or down via a new control on claude.ai and Cowork. In Claude Code, the dynamic workflows feature enables the model to plan and execute hundreds of parallel subagents in a single session, handling codebase-scale migrations end-to-end. Pricing remains unchanged from Opus 4.7 ($5/$25 per million input/output tokens), while fast mode (2.5× speed) is now three times cheaper. Benchmarks show strong gains: Opus 4.8 is the only model to complete all cases on the Super-Agent benchmark, scores 84% on Online-Mind2Web for browser agents, and sets records on legal and financial-domain evaluations.

The takeaway for builders is that Opus 4.8 is a modest but tangible upgrade—sharper judgment, better tool-use efficiency, and proactive honesty. The new effort control gives engineers fine-grained tradeoffs between latency and quality, while dynamic workflows unlock large-scale autonomous coding. Anthropic also teases upcoming Mythos-class models under Project Glasswing, which will offer even higher intelligence but require stronger cyber safeguards before broad release. Serious teams should evaluate Opus 4.8 for agentic and high-stakes workflows where reliability matters most.