Anthropic today announced Claude Opus 4.8, which has improvements in all benchmarks, and is a very efficient collaborator.
Opus 4.8 is introduced along with several new features. Users on claude.ai now have control over the amount of effort Claude puts into a job. Claude Code has a new “workflow” feature that allows it to handle even larger problems. And Opus 4.8’s fast mode—where the model can run at 2.5× speed—is now three times cheaper than previous models.
Power of Opus 4.8
One of the most notable improvements in Opus 4.8 is its reliability, for example, to avoid making claims they cannot support. But a common problem with AI models is that they sometimes jump to conclusions, confidently claiming to have made improvements in their work despite little evidence. Early testers report that Opus 4.8 may flag uncertainty about its functionality and is less likely to make unsupported claims. This is proven in the analysis, which shows that Opus 4.8 is almost four times smaller than its predecessor to allow errors in the code it wrote to pass unnoticed.
A detailed alignment test was performed on the model before release by Anthropic’s Alignment team, and the company concluded that Opus 4.8 “reaches a new high in our measurements of promotional features such as supporting user autonomy and acting in a way that benefits the user.” Tests also showed Opus 4.8 to have levels of bad behavior (such as cheating or collusion and abuse) that are much lower than Opus 4.7, and similar to our best-aligned model, Claude Mythos Preview. Full alignment testing, along with a list of pre-shipment security checks, is reported on the Claude Opus 4.8 System Card.
In addition to Claude Opus 4.8, the following updates have also been made:
- Powerful workflow. Available in research preview, this feature allows Claude to perform even greater tasks in Claude Code. Claude can schedule a job and run hundreds of the same subagents in a single session (and with Opus 4.8, agents can run even longer). It then validates the output before reporting it to the user. For example, Claude Code with Opus 4.8 can now perform codebase-scale migrations of hundreds of thousands of lines of code from kickoff to integration, with an existing test suite as its backbone. You can learn more about flexible workflows—available in Claude Code for Enterprise, Team, and Max editions—in this post.
- Effort control at claude.ai and Cowork. A new control along with the model selector allows users to choose how much effort Claude puts into the response. In higher effort settings, Claude will think more often and more deeply to provide better answers. At low effort settings, Claude will respond quickly and use user rate limits very sparingly. Users now have this option—effort control is available for all programs.
- The Messages API now accepts system entries within the message list. Developers can update Claude’s commands in the middle of a job without breaking the data cache or submitting an update to the user’s rotation. This can be used in a given harness to update permissions, token budgets, or local context as the agent runs.
Claude Opus 4.8 is available everywhere today. The price for general use has not changed from Opus 4.7: $5 per million input tokens and $25 per million output tokens. The price of fast mode is $10 per million input tokens and $50 per million output tokens. Developers can use claude-opus-4-8 with the Claude API.
