Anthropic ships Claude Opus 4.8 with 1890 on GDPval-AA, same price
Adds a five-level Thinking selector, research-preview fast mode at 2.5x speed, and dynamic workflows that spin up hundreds of subagents in Claude Code.
By Ryan Merket · Published
Why it matters
Opus 4.8 pushes state-of-the-art benchmark performance while keeping price flat, and adds speed and orchestration features that target real software delivery, not just chat UX.
Anthropic's new Claude Opus 4.8 scored 1890 on the GDPval-AA agentic real-world benchmark, 137 points above Opus 4.7 and 121 points ahead of the next-best model, Anthropic (@claudeai) said in a thread on X. The model is available at the same price and introduces a five-level Thinking effort selector and improved autonomous work capability, with general availability detailed in an Anthropic news post.
https://x.com/claudeai/status/2060042702150930686
Opus 4.8 is live on the web, the Claude Platform, and major cloud providers, according to the announcement. In Claude Code, Anthropic says Opus 4.8 can stay on track across long-running sessions and follow work through a repo, enabling handoffs for things like feature builds or bug sweeps.
Anthropic also introduced a fast mode for Opus 4.8. It is the same model running roughly 2.5x faster and priced at three times cheaper than before, the company wrote. Fast mode is available as a limited research preview; developers can toggle it with /fast in Claude Code or request API access via the fast mode page.
For complex engineering tasks, Anthropic unveiled dynamic workflows in Claude Code as a research preview. The system plans work, runs tens to hundreds of parallel subagents, and verifies results before reporting back, aimed at jobs like multi-file migrations. Details are in the product blog post.