Anthropic releases Claude Opus 4.8 with coding, agent upgrades

Anthropic released Claude Opus 4.8, an update to its Opus series. The model is available through claude.ai, Claude Code and the Claude API under the identifier claude-opus-4-8.

Anthropic positioned Opus 4.8 for coding and agentic tasks. The company reported improvements over Opus 4.7 on benchmarks for coding, agent skills, reasoning and office productivity. The model can use tools within a context and perform self-checks on its outputs.

Anthropic reported Opus 4.8 is four times less likely to pass flawed code without comment compared with Opus 4.7. The update also showed lower rates of deceptive or permissive behavior, with behavior reported as comparable to Claude Mythos Preview.

The release adds an effort control setting that changes how much computation the model applies to each response and therefore how many tokens it consumes. The default level is “high”; users can select “xhigh” for greater computation and higher token use. On coding tasks Anthropic reported the higher default uses about the same number of tokens as Opus 4.7 while delivering better performance. Claude Code rate limits have been raised to support increased token use.

Claude Code now supports dynamic workflows for large codebases. Workflows can plan tasks, run parallel sub-agents, verify outputs and report results. Anthropic noted the feature can migrate codebases of hundreds of thousands of lines. Dynamic workflows are in research preview and available to Enterprise, Team and Max customers.

The Messages API accepts live edits to the messages array. Developers can update instructions, permissions or token budgets while an agent continues working without breaking prompt cache or requiring a new user turn.

Anthropic published pricing tied to token use. Outside of “fast” mode, input tokens cost $5 per million and output tokens cost $25 per million. Fast mode doubles those prices to $10 per million input and $50 per million output and runs at roughly 2.5 times the speed. The company noted it is shifting from subscription tiers to token-based billing to make cost and effort trade-offs explicit to users.

Organizations in software development, law, finance and research tested Opus 4.8 before wider release. One tester reported cost parity with GPT-5.5 on internal benchmarks. Benchmarker CursorBench reported Opus 4.8 required fewer tool steps to reach comparable outputs. Anthropic published a System Card with additional detail on model behavior and recommended customers review it.

Anthropic described a roadmap that includes developing lower-cost models with similar capabilities and more advanced Mythos-class models. The company expects Mythos-class models will require stronger safeguards and plans to make them available to customers in the coming weeks. Project Glasswing was cited as an example of early enterprise testing of Claude Mythos Preview for cybersecurity scanning.