Anthropic has released Claude Opus 4.8, an upgrade to Claude Opus 4.7. The company says this improves results for coding, agent work, reasoning, and knowledge work. The platform is available through claude.ai, Claude Code, and the Claude API with the API name claude-opus-4-8.
The company also changed some details of its product lineup. Claude.ai and Cowork users can configure the amount of effort Claude applies to responses. This basically affects how many tokens the model burns. Claude Code also has dynamic workflows, the ability to plan work, run subagents in parallel, validate output, and report back to the user. Finally, the Message API accepts live changes to message arrays. Anthropic says this allows developers to update instructions mid-task without interrupting prompt caching or requiring additional user interaction.
According to Anthropic, Claude Opus 4.8 will still cost $5 per million input tokens and $25 per million output tokens when not in “fast” mode, while in fast mode it will cost $10 per million input tokens and $50 per million output tokens. Opus 4.8’s high-speed mode runs at 2.5x speed, the company said after the announcement.
The company is positioning Opus 4.8 as designed for agent workflows in coding and coding, allowing models to use tools and see their work in context. Opus 4.8 improves on Opus 4.7 in coding, agent skills, reasoning, and administrative benchmarks. There are system cards that allow you to examine even more subjective details.
Anthropic’s announcement lists several companies that have tested the platform ahead of its wider release, including companies operating in the fields of software development, law, finance, and research. Several testers have commented on the platform’s agent workflow, with one tester noting that the cost of running internal benchmark tests is comparable to GPT-5.5. According to comments on CursorBench, Opus 4.8 uses fewer steps in the tool to achieve the same level of output.
Anthropic says Opus 4.8 is less likely to pass defective code without comments than its predecessor, 4.7, by one in four. The platform states that it has a lower deception rate and propensity for abuse than Opus 4.7, and is comparable in this respect to what was demonstrated in the Claude Mythos Preview.
Effort control allows users to manage trade-offs between quality, speed, and token burn rate. Opus 4.8’s defaults are high effort, but the company says that for coding tasks, the higher defaults use only Opus 4.7 type token numbers, but provide better performance. Users can select “xhigh” for tasks that require more calculations. Anthropic said it has increased the rate limit on Claude Code to support the resulting increase in token usage.
Claude Code’s dynamic workflows are designed for large codebases and can migrate codebases of hundreds of thousands of lines. These features are currently in research preview and available on Enterprise, Team, and Max plans.
The Message API updates instructions while the agent is running. For example, edits in the message array are used to update permissions or change a token’s budget or context while the agent continues working.
Anthropic also hinted that it is using this release to develop a model that provides users with current levels of functionality at a lower cost, and that it plans to release a model that is a better class than the current Opus platform. That roadmap includes Project Glasswing, under which a group of organizations is using Claude Mythos Preview for cybersecurity scanning. Anthropic said its feature-level models require stronger safeguards before being released to all customers. The company plans to offer “Mythos class” models to customers in the coming weeks.
The additional controls in 4.8 make cost and effort trade-offs clearer to users as companies move from subscription tiers to token-based billing.
(Image source: Pixabay, under license.)
Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expos in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other major technology events. Click here for more information.
AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.
