Anthropic launches Claude Opus 4.8 with coding and agent upgrades

Anthropic has released Claude Opus 4.8, an update to Opus 4.7 available through claude.ai, Claude Code and the Claude API (api name: claude-opus-4-8). The company positions the release for coding and agentic workflows that use tools, check outputs and work across large contexts.

Anthropic reported improvements on benchmarks for coding, agent skills, reasoning and office tasks compared with Opus 4.7. The company also said Opus 4.8 is four times less likely than Opus 4.7 to return flawed code without comment and shows lower rates of deceptive or misuse-prone behavior, with performance the company describes as comparable to its Claude Mythos Preview on those measures.

The update introduces user controls for computational effort, which affect the number of tokens a response consumes. Effort defaults to a “high” setting; Anthropic said that on coding tasks the higher default consumes roughly the same number of tokens as Opus 4.7 while delivering better results. An “xhigh” option is available for tasks that need more computation. Pricing remains token-based: non-fast mode is $5 per million input tokens and $25 per million output tokens. Fast mode runs about 2.5 times faster and costs $10 per million input and $50 per million output tokens.

Claude Code gains “dynamic workflows” that plan work, run parallel sub-agents, verify outputs and return consolidated results. Anthropic described the feature as suitable for very large codebases and capable of migrating repositories of hundreds of thousands of lines. Dynamic workflows and related capabilities are available in research preview on Enterprise, Team and Max plans, and the company increased Claude Code rate limits to support higher token use.

The Messages API now accepts live edits to the messages array so developers can change instructions, permissions, token budgets or context while an agent is running, without breaking prompt cache benefits or requiring a separate user turn. The announcement frames the change as a way to keep agents operating through changing constraints while preserving efficiency.

Anthropic cited early tester feedback from companies in software development, law, finance and research. One tester reported cost parity with GPT-5.5 on its internal benchmarks. A benchmarking participant named CursorBench noted that Opus 4.8 used fewer tool steps to reach the same output quality compared with prior models.

The company said a System Card for Opus 4.8 is available for more subjective evaluation and safety details. Anthropic also signaled ongoing model development beyond the Opus line, saying it is working on models that aim to deliver current capability at lower cost and expects to make “Mythos-class” models available to customers in the coming weeks, pending stronger safeguards. Under Project Glasswing, a group of organizations is using Claude Mythos Preview for cybersecurity scanning.

Anthropic added that the new controls will expose cost and effort trade-offs as it moves more of its product line to token-based billing.