Anthropic released Claude Opus 4.8 on May 28 with three practical additions: an effort control setting, parallel subagent workflows inside Claudee Code](/tools/claude-code/), and a significant price cut on fast mode inference.
Base API pricing holds at $5 per million input tokens and $25 per million output tokens - same as Opus 4.7. Fast mode, which trades some output quality for speed, drops to $10 and $50 per million tokens respectively. That's roughly three times cheaper than the previous fast mode rates.
Effort control lets you set how hard the model works on a given response. Higher settings trigger more deliberate, multi-step reasoning; lower settings prioritize speed. For developers building agentic pipelines - automated workflows where the model handles a sequence of tasks without human input - this means you don't have to run full reasoning on every low-stakes subtask. Claude Code users also get dynamic workflows, which spin up parallel subagents for large codebase work. Anthropic described the target use case as tasks spanning hundreds of thousands of lines of code. A third addition to the Messages API now allows system-level instructions mid-conversation, so developers can update task parameters without restarting a session from scratch.
Four Times Fewer Missed Code Flaws
The reliability improvement Anthropic highlights most: Opus 4.8 is roughly four times less likely than 4.7 to let a code flaw pass without flagging it. That specific failure mode - a model that silently accepts broken code rather than pushing back - made earlier versions unreliable for unsupervised coding sessions.
On benchmarks, the model hits 84% on Online-Mind2Web, which tests how accurately a model can control a web browser to complete real tasks. On the Super-Agent benchmark - a suite of complex, multi-step agentic problems - it completed every case and outperformed GPT-5.5 at comparable cost. It also posted the highest recorded score on the Legal Agent Benchmark.
Anthropic reports the model hits "new highs" on what they call prosocial traits: correctly declining harmful requests and catching its own mistakes. These behavioral claims are harder to verify independently outside controlled evaluations, but the directional trend across the 4.x series has been consistent.
Opus 4.8 is available now on claude.ai and via the API using the model ID claude-opus-4-8.