What Is Claude Opus 5? The New Effort Setting, Pricing and Who It Is For

In short: Anthropic released Claude Opus 5 on 24 July 2026. The headline change is not model size but that you can now choose how much effort the model spends on a task. Pricing stayed the same as Opus 4.8 while capability rose sharply.

What is Opus 5?

Claude Opus 5 is Anthropic's new flagship in the Opus class, the series reserved for demanding work. The ladder runs like this: Haiku for speed and low cost, Sonnet for the balanced load of daily work, Opus for demanding tasks, and Mythos as the research class at the very top. What makes Opus 5 notable is that it shifts that balance: it approaches the capability of Fable 5, Anthropic's strongest model, at half the cost.

In practice this means most people and companies who could not afford the strongest model can now use something very close to it. That has been the clearest trend in AI tools for two years running: the same capability arrives cheaper with every release.

The big change: the effort setting

The most concrete addition in Opus 5 is the effort setting. You tell the model how deeply to think about a task: low, medium or high. The API adds a tier above that.

Why does this matter? Because not everything you do with AI carries the same weight. Fixing an email and analyzing a contract do not require the same mental load. Until now the model approached both with the same depth, which meant unnecessary waiting and unnecessary cost on simple work.

• Low effort: Short answers, formatting, translation, summarizing. Fast and cheap.
• Medium effort: Most daily work. Writing, analysis, planning.
• High effort: Complex analysis, long running code tasks, decisions that need multi step reasoning.

Speed and cost sit on one side, depth on the other. You now set that balance according to the nature of the work. The free prompt library shows which jobs carry which weight, which makes the choice easier.

Fast mode: what you pay for speed

The second new option is Fast mode. The model runs at roughly 2.5 times its default speed and you pay twice the base price for it. That is worth it in a live customer conversation or a demo where someone is waiting in front of you. For a batch job running overnight it is money wasted.

Technical specifications

A 1 million token context window means you can hand the model roughly 700 thousand words at once. That is several times a mid sized book. A year of customer correspondence, an entire set of contracts, or every file in a large project can go into a single conversation.

What the benchmarks say

These are the results Anthropic published:

Read benchmark numbers carefully. These tests measure performance on specific task types and they are the vendor's own measurements. They do not tell you how much difference the model makes in your work. They do give a signal though: there is a clear jump on multi step, tool using work, which is what OSWorld and AutomationBench measure.

The real change: the model checks its own work

The change that will matter more to you than any number is this: Opus 5 verifies its own work and recovers from its own mistakes. Anthropic presents this as the model's defining trait.

In daily use that means less back and forth. With earlier models, getting something right usually took three or four rounds: the model produces something, you point out what is missing, it fixes that, something else breaks. Opus 5 handles some of those rounds internally.

That does not mean you should stop checking. The "audit the AI's own work" prompt in the library exists exactly for this: asking the model to critique its own output as if someone else wrote it lowers the error rate noticeably.

Who is it for?

• Heavy users: If you work with AI for hours a day, the effort setting saves you time and money directly. Putting simple work on low effort and heavy work on high effort means more done on the same budget.
• People working with long documents: Lawyers, accountants, consultants, researchers. A 1 million token window changes the rules for these professions.
• People writing or commissioning code: The CursorBench result and the self verification behavior show up most clearly in long coding tasks. That matters especially if you build apps through vibe coding.
• Scientific researchers: Anthropic highlights biology and scientific research tasks in particular.

If you only ask the occasional question and use AI a few hours a month, Opus 5 changes nothing for you. The model on the free tier is already more than enough for most everyday work.

Which Claude plan has it?

If you want to understand how the Opus class sits against Anthropic's top tier, see the Fable 5 and Mythos 5 guide. Learning the tool makes more sense than picking a model first.

Limits and things to watch

• Fable 5 is still recommended for the hardest autonomous work. Anthropic continues to point at Fable 5 for long running, fully autonomous projects.
• Mythos 5 leads on cybersecurity. Opus 5 also ships with tighter guardrails in that area.
• Pricing is the same but usage can rise. The per million token price matches Opus 4.8. But on high effort the model thinks more, and thinking spends tokens. Your bill tracks the weight of the work.
• The benchmark numbers are the vendor's own. Independent tests arrive over time, so test it on your own work before deciding.

How to get started

1. Sign in at claude.ai. On Pro or Max, Opus 5 is already in the model list.
2. Introduce yourself in your first message. Say what you do and how you want to be answered. This single step visibly improves answer quality.
3. Make it a habit to pick low effort for simple work and high effort for heavy work.
4. If you are working with long documents, give them all in one conversation. That is what the 1 million token window is for.
5. Do not use the output as is. Have the model audit its own work.

There is a ready made prompt for every one of these steps in the prompt library, free to download.

Conclusion

What Opus 5 brings is control more than raw power. Being able to choose how much the model thinks and how fast it runs is a more valuable change than a version number for anyone who has actually put AI into their workflow. Capability rising while the price holds steady also points at the direction: these tools get a little cheaper and a little more useful every month.

The real question is not which model you use but how you fit it into your workflow. That is exactly what my AI training covers.