Claude Opus 5: The Interesting Part Isn't the Benchmarks

Anthropic released Claude Opus 5 on July 24. The positioning is unusually plain for a model launch: an everyday model that, per Anthropic, comes close to Claude Fable 5's frontier intelligence at half the price. The pricing backs that up. Opus 5 costs $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8. The capability moved; the invoice did not.

The API model ID is `claude-opus-5`. It is available now on all platforms and all paid Claude plans. It becomes the new default on Claude Max and the strongest model available on Claude Pro. Anthropic's docs recommend it for complex agentic coding and enterprise work.

Those are the launch facts. The more useful question for anyone running coding agents is what kind of model this is, and that is where the benchmark numbers are the least interesting part of the story.

The reported numbers, briefly

Anthropic reports state-of-the-art results on coding and knowledge-work evaluations, including Frontier-Bench and GDPval-AA. On CursorBench 3.2 at max effort, the company reports performance within 0.5% of Fable 5's peak at half the cost per task. It also reports an ARC-AGI 3 score three times the next-best model, and says Opus 5 more than doubles Opus 4.8 on Frontier-Bench while costing less per task.

Every one of those figures comes from Anthropic. None have independent replication yet. Treat them as the vendor's claims about the vendor's model, because that is what they are. The honest reading is directional: Anthropic believes it has pushed near-frontier capability into an everyday price point, and it published numbers consistent with that belief.

Why single-task benchmarks undersell agentic work

Here is the argument I think matters. A benchmark measures one attempt at one task. An agent session is dozens or hundreds of attempts chained together, where each step consumes the previous step's output. In that setting, error handling dominates error rate.

Suppose a model solves individual steps correctly 95% of the time and never checks its work. Over a 30-step task, mistakes accumulate and the session quietly derails, usually in a way you discover an hour later. Now suppose a model solves steps at a similar rate but verifies its output, notices when something is off, and repairs it before moving on. The second model finishes tasks the first one cannot, even if their single-shot scores look identical. Verification compounds. Raw accuracy does not.

This is why the behavioral claims in the Opus 5 launch are more consequential than the leaderboard ones. Anthropic describes the model as one that verifies its work, iterates carefully, and is stronger on long-running multi-step tasks. Two specific behaviors stand out:

**Root-cause diagnosis.** Per Anthropic's description, Opus 5 tends to look for the underlying cause of a failure before changing code, rather than patching the symptom. Anyone who has watched an agent "fix" a failing test by editing the assertion knows why