Anthropic releases Claude Opus 5

Anthropic released Claude Opus 5 on July 24, 2026. The company positions it as coming close to the frontier intelligence of Claude Fable 5, its most capable model, at half the price. It is now the default model on Claude Max and the strongest model offered on Claude Pro.

Pricing stayed at $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8. That is the quiet headline. A generational upgrade normally comes with a bill increase, and this one did not.

What you actually get for the money

A token is roughly three quarters of a word, so a million input tokens is about 750,000 words of material you send the model, and a million output tokens is about 750,000 words it writes back. Output is the expensive side at five times the input rate, which is why a model that reaches the same answer in fewer words is cheaper in practice even at the same sticker price.

There is also a fast mode that runs about 2.5 times quicker for twice the base price. That trade makes sense for anything a person is sitting and waiting on, and much less sense for a batch job running overnight.

You can reach it in the API as claude-opus-5, and it is live in Claude.ai, Claude Code and Claude Cowork.

How it compares, in plain terms

Anthropic's benchmark selection is heavily weighted toward agents, meaning software that goes off and completes a multi step task rather than answering one question. Here is what the tests actually measure.

Against Opus 4.8, the model it replaces:

• Frontier-Bench v0.1, a broad capability test: Anthropic reports Opus 5 roughly doubling Opus 4.8's score while costing less to run.

Against Fable 5, the expensive flagship:

• CursorBench 3.2, which measures real coding work inside an editor: Opus 5 lands within 0.5% of Fable 5, at half the cost. Statistically that is a tie for a fraction of the money.
• OSWorld 2.0, which tests whether a model can operate a computer, clicking through actual applications to finish a task: Anthropic says Opus 5 beats Fable 5's best result at one third the cost.

Against everything else:

• ARC-AGI 3, a set of visual reasoning puzzles designed to be hard for models and easy for people: Anthropic claims a score three times higher than the next best model. This is the boldest claim in the release and the one most likely to draw scrutiny.
• It also tops the Zapier AutomationBench leaderboard, which scores models on stringing real business tools together.

Anthropic additionally points to better verification behavior, meaning the model checks its own work before handing it back, plus improved visual output and stronger results on scientific tasks. Organic chemistry and protein structure prediction were named specifically.

The safety picture

Anthropic calls Opus 5 its most aligned model so far, with the lowest misaligned behavior scores it has recorded. Alignment here is the practical question of whether the model does what was asked rather than something adjacent to it, which matters a great deal more when the model is running unsupervised for an hour.

Two details sit underneath that. The cyber classifiers, the filters that decide whether a security question gets answered, are 85% less restrictive than Fable 5's. So security professionals should hit fewer unnecessary refusals. And Opus 5 deliberately remains behind Claude Mythos 5 on offensive cybersecurity tasks. Mythos 5 is the restricted model the US cleared for trusted partners in June, and Anthropic is keeping that capability gap between its public and restricted lines on purpose.

What is still unknown

No independent evaluation existed at launch, so treat the numbers above as vendor claims until outside researchers reproduce them. Anthropic also did not publish a context window figure, which is the amount of material the model can consider at once, and said nothing about whether fast mode costs you any quality. Those answers usually arrive within a week or two once people run their own tests.