Anthropic released Claude Opus 5 on 24 July 2026, two months after Opus 4.8. Most coverage led with a price line: frontier intelligence at half the cost. That line is accurate against one model and misleading against another. This piece covers what shipped, what it costs, where it is available and what Anthropic claims about performance. How to run the model, including effort settings and prompt changes, is in our operating notes.

What shipped

Opus 5 has a 1M-token context window, which is both the default and the maximum, and 128K maximum output tokens. According to Anthropic’s documentation, thinking is on by default: the model decides when and how much to reason on each turn, and an effort parameter running from low through medium, high, xhigh and max controls how deep that goes.

It is available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. In the consumer apps it became the default on Claude Max and the strongest model on Claude Pro. Opus 4.8 stays available everywhere it was before.

Two smaller changes matter more than they look. The minimum cacheable prompt dropped from 1,024 tokens to 512, so short prompts that could not be cached before now can be, with no code change. And batch processing takes 50% off both rates, for work that does not need an immediate answer.

The price is not a cut

Opus 5 costs $5 per million input tokens and $25 per million output tokens. That is the same as Opus 4.8, and the documentation says the pricing is unchanged. The half-price headline compares it with Fable 5, which launched in June at $10 and $50. Anthropic presents Opus 5 as similar performance at half of Fable 5’s rate. That is a real claim, and it is not a price cut: if you budgeted on your existing Opus spending halving, it will not.

Your bill falls only if you were paying Fable rates for work that Opus 5 can now do, or if you move suitable work to batch processing.

What Anthropic says about performance

Anthropic’s announcement makes three comparative claims. On OSWorld 2.0, a computer-use benchmark, Opus 5 surpasses Fable 5’s best result at just over a third of the cost. On Frontier-Bench v0.1 it more than doubles Opus 4.8’s performance at a lower cost per task. On ARC-AGI 3, a test aimed at novel reasoning, its score is three times that of the next-best model.

These are the vendor’s figures at launch, and we have not reproduced them. A benchmark score is also not a promise about your workload: a model that tops an agentic test will not necessarily write better product descriptions for your shop. The number that matters is the one you measure on your own tasks.

Does it need action?

For most people using Claude through an app, no. The model behind the tool improved and the price of the tool did not move. For anyone on the API, the specifications above are the changes to check against your setup, and the effort, verification and injection questions are in the operating notes linked at the top.

Sources

Performance figures are as stated by Anthropic at launch. We have not independently reproduced them.