原文 · 未翻译
Anthropic has released Claude Opus 4.8, an upgrade to Claude Opus 4.7 that the company says brings improved results for coding, agent work, reasoning, and knowledge work. The platform can be used through claude.ai, Claude Code and the Claude API, with the API name claude-opus-4-8.
The company has also altered some of the details of its product line-up. Users of claude.ai and Cowork can set the amount of effort Claude applies to a response – essentially, affecting the number of tokens the model will burn. Claude Code also has dynamic workflows, a feature that plans work, runs parallel sub-agents, verifies outputs and reports back to the user. Finally, the Messages API accepts live changes to the messages array, which Anthropic says lets developers update instructions during a task without breaking prompt cache use or needing a separate user turn.
Anthropic said the price for use of Claude Opus 4.8 when not in ‘fast’ mode will remain at $5 per million input and $25 per million output tokens, while fast mode costs $10 per million input tokens and $50 per million output. Fast mode for Opus 4.8 works at 2.5x, the company’s announcement post states.
The company has positioned Opus 4.8 as designed for coding and agentic workflows in coding, where the model can use tools inside a context and check its own work. It says Opus 4.8 improves on Opus 4.7 on benchmarks for coding, agent skills, reasoning, and office work. There is a System Card that can be examined for further subjective detail.
Anthropic’s announcement cites several companies that have tested the platform before its wider release, including those operating in software development, law, finance, and research. Several testers commented on the platform’s agentic workflows, with one noting a cost parity with GPT-5.5 when running its internal benchmark tests. A comment from CursorBench said Opus 4.8 used fewer tool steps to achieve the same level of output.
Anthropic says Opus 4.8 is less likely than its 4.7 predecessor to pass flawed code without comment, which it describes four times less likely. It says the platform showed lower rates of deception or the tendency to go along with misuse than Opus 4.7 and is comparable in this regard as those exhibited by Claude Mythos Preview.
Effort control helps users to manage any trade-off between quality, speed, and token burn rates. Opus 4.8 defaults to high effort but on coding tasks, the company said the higher default only uses the type of token numbers of Opus 4.7, but performs better. Users can opt for ‘xhigh’ for work that needs more computation. Anthropic said it has increased Claude Code rate limits to support the resulting higher token use.
Dynamic workflows in Claude Code are designed for large codebases, and can migrate codebases of hundreds of thousands of lines. These features are currently in research preview and are available on the Enterprise, Team, and Max plans.