Anthropic launches Claude Opus 5 at the same price

Anthropic now offers Claude Opus 5, a new model designed to approach the performance of its most advanced model, Claude Fable 5, but at half the cost. It also replaces Opus 4.8 without raising the price.

Opus 5 is available today through the Claude API, Claude Max, Claude Pro and other platforms. Anthropic presents it as a model built for everyday use, capable of working for longer, checking its own results and completing tasks with less human intervention.

More performance without paying more

The base price remains 5 dollars per million input tokens and 25 dollars per million output tokens, the same rates as Opus 4.8. Tokens are the units of text the model processes.

In Anthropic's published tests, Opus 5 outperforms other models in programming, analysis and office tasks. In Frontier-Bench, a software engineering evaluation, it more than doubles the performance of Opus 4.8 at a lower cost per task.

In CursorBench, a programming-focused test, it achieves a score less than 0.5% below Fable 5's maximum when running at the highest effort level, but at half the cost per task.

The model also stands out in tasks that go beyond writing code:

• In ARC-AGI 3, it scores three times higher than the next model.
• In Zapier AutomationBench, it completes end-to-end business tasks with a success rate approximately 1.5 times higher than the next model at the same cost.
• In OSWorld 2.0, which measures computer use, it outperforms the other models at every cost level and exceeds Fable 5's best result while spending just over one third as much.

These figures come from Anthropic's evaluations, so they provide an indication of performance but do not replace independent testing in real-world situations.

A model that checks its own work

The main improvement is not just that it responds better, but that it works more carefully. Anthropic says Opus 5 verifies its results, detects errors during planning and changes strategy when something goes wrong.

In one test, it received a drawing of a mechanical part and had to reconstruct it as a 3D model without being able to view the image directly. The model created its own process for extracting the geometry from the pixels and completed the task repeatedly.

In another case, it found the cause of an error in an open-source code package manager and fixed a problem that an earlier patch had not resolved. It also built its own testing system to validate a market data reader when there was no live source available for comparison.

For you, this means it can take on longer assignments, such as reviewing a codebase, building a complete application, analyzing financial documents or preparing a presentation and correcting it afterward. Human oversight is still necessary, especially when a task has legal, financial or security implications.

Improvements in science and business work

Opus 5 also improves on Opus 4.8 in life sciences evaluations. Anthropic reports an advantage of 10.2 percentage points in organic chemistry tasks and 7.7 points in protein-related predictions.

Comments from users with early access mention improvements in financial analysis, research, legal review, business automation and interface design. Box, for example, says the model outperforms Opus 4.8 by 8% in its tests, with improvements of 11% in data analysis and 17% in due diligence processes.

Opus 5 also lets you adjust its effort level. At a low level, it responds faster and uses fewer tokens. At higher levels, it dedicates more resources to reasoning and reviewing the result. Anthropic says it can maintain performance similar to Opus 4.8 while using 26% fewer tokens on certain legal tasks.

Limits and safety measures

The model is not the most capable in every area. Anthropic says it still trails Mythos 5 in cybersecurity, especially when turning a discovered vulnerability into a functional attack. However, it comes close to that model when locating vulnerabilities.

Opus 5 can analyze code flaws, but it blocks tasks such as binary-based vulnerability scanning, penetration testing and exploit generation. Anthropic has also added options to automatically redirect blocked requests to another model instead of failing without a response.

Claude Opus 5 can now be used with the identifier claude-opus-5. There is also a fast mode that runs approximately 2.5 times faster, although it doubles the base price. The important question now is not whether it can beat a specific test, but how much reliable work it can complete from start to finish when given access to tools and real-world context.