Anthropic launches Claude Opus 4.8 for complex tasks

Anthropic has launched Claude Opus 4.8, a new version of its most advanced model, with improvements in coding, analysis, tool use and tasks that require multiple steps. It is already available to all users and keeps the same price as Opus 4.7.

The update is not focused only on giving better answers. It also aims to make Claude a more reliable collaborator: one that detects errors, questions weak plans and recognizes when it does not have enough information.

What improves in Claude Opus 4.8

Anthropic describes Opus 4.8 as a moderate but noticeable improvement over its predecessor. According to its evaluations and early users, the model performs particularly well in:

• Coding and maintaining large codebases.
• Research, analysis and work with complex documents.
• Using browsers and applications through tools.
• Legal, financial and knowledge-management tasks.
• Long-running work where it must retain context and follow instructions for longer periods.

One of the clearest figures is that the model is about four times less likely than Opus 4.7 to overlook errors in the code it has written without flagging them. Anthropic also says it is better at detecting problems in input data and in its own output.

That does not mean it is infallible. It means that, in theory, it should warn you more often when an answer needs review instead of presenting a questionable conclusion as if it were certain.

In a computer and browser-use evaluation, Opus 4.8 scored 84% on Online-Mind2Web, according to Anthropic. The company also cites results from external collaborators that place it above earlier models on agent tasks, legal analysis and coding. These figures come from specific tests and early-user reports, and do not guarantee the same performance in every real-world situation.

Three changes you will notice

The update comes with several new features.

Effort control

In claude.ai and Cowork, a control lets you decide how much time and how many resources Claude dedicates to a task. At a low level, it responds faster and uses less of your usage limit. At higher levels, it analyzes more deeply and may produce better results on difficult work.

Opus 4.8 uses the high level by default. It also offers the extra and max modes, designed for complex tasks or processes that can run for a long time. In Claude Code, the extra level is identified as xhigh.

Dynamic workflows in Claude Code

The dynamic workflows feature, available as a research preview, can divide a large job among hundreds of subagents running in parallel. A subagent is a specialized instance that handles a specific part of the task.

For example, Claude Code can attempt to migrate a codebase with hundreds of thousands of lines, check the result using existing tests and continue until it is ready for integration. The feature is available on Enterprise, Team and Max plans.

Lower cost for fast mode

Fast mode lets you work with Opus 4.8 at 2.5 times its usual speed. Anthropic says it now costs three times less than the equivalent mode for its earlier models.

In the API, standard usage costs $5 per million input tokens and $25 per million output tokens. Fast mode costs $10 and $50, respectively. A token is a unit of text the model processes, not exactly a word.

What this means for you

If you use Claude to write, research or code, the main change should be greater consistency during long sessions. The model tries to retain context more effectively, use fewer steps when calling tools and warn you sooner when something does not add up.

For businesses, the difference may be more practical: an agent that reviews its work and flags its uncertainties needs less supervision than one that delivers confident but incorrect results. Even so, its output still needs to be verified when dealing with code, legal matters or financial decisions.

Developers can already use the model through the API with the identifier claude-opus-4-8. Anthropic also allows system instructions to be inserted into the message list, making it easier to update an agent's permissions, token budgets or context while it works.

The company is also preparing cheaper models with capabilities similar to Opus, along with a future Mythos family with greater capacity. For now, Mythos is being tested by a small group of organizations on cybersecurity tasks. Its general release will depend on Anthropic completing additional safety measures.

For now, Opus 4.8 represents an incremental improvement with a clear direction: fewer confidently unsupported answers, more autonomy on long-running work and greater control over how much effort the model applies. The important question will be whether those improvements hold up outside testing and in day-to-day work.