Anthropic · 2026-05-28 · seismic
Claude Opus 4.8
Anthropic releases Claude Opus 4.8 with sharper agentic judgment, ~4x fewer unremarked code flaws, a dynamic-workflow research preview for running hundreds of parallel subagents, and an effort-control option on claude.ai. Pricing unchanged from Opus 4.7.
Anthropic's flagship gets sharper agentic judgment, honest error-flagging, and a parallel-subagent dynamic-workflow tool.
Key specs
| Input per million usd | 5 |
|---|---|
| Output per million usd | 25 |
| Fast mode input per million usd | 10 |
| Fast mode output per million usd | 50 |
| Online mind2web pct | 84 |
| Agentic coding pct | 69.2 |
What is it?
Claude Opus 4.8 is Anthropic's latest flagship model, released 41 days after Opus 4.7. It ships alongside a dynamic-workflow research preview that can orchestrate hundreds of parallel subagents, plus an effort-control option on claude.ai. Pricing for standard usage is unchanged at $5/$25 per million tokens.
How does it work?
Opus 4.8 is trained to flag its own errors ~4x more reliably than Opus 4.7, reducing silent code flaws and unsupported analytical claims. The dynamic-workflow tool is a separate orchestration layer that distributes tasks to parallel Claude subagents, enabling codebase-scale migrations across hundreds of thousands of lines. An effort slider trades thinking depth for 2.5x speed via fast mode.
Why does it matter?
More honest error-flagging directly reduces debugging cost in AI-assisted development. The dynamic-workflow feature moves from sequential agent chains to true parallelism, cutting wall-clock time on large refactors. Same pricing as Opus 4.7 means teams can upgrade without renegotiating budgets.
Who is it for?
Software engineers and analysts running long-horizon agentic tasks.
Try it
API model ID: claude-opus-4-8