Anthropic released Claude Opus 5 on 24 July 2026, and the headline is as much about the price tag as the capability: it ships at $5 per million input tokens and $25 per million output tokens — identical to Claude Opus 4.8, the model it succeeds. A frontier upgrade that does not raise the per-token price is unusual enough to be the story on its own.
The claims
Anthropic's announcement is heavy on comparisons and light on raw numbers. The company reports results as charts and relative statements rather than a benchmark table, so there are no absolute percentages to quote. What it does claim:
- Frontier-Bench v0.1 — Opus 5 "surpasses all other models, and more than doubles Opus 4.8's performance."
- ARC-AGI 3 — its score is "three times as high as the next-best model."
- CursorBench 3.2 — lands "within 0.5% of Fable 5's peak score, but at half the cost per task."
- OSWorld 2.0 — surpasses Fable 5's best result "at just over a third of the cost."
- Zapier AutomationBench — pass rate is "around 1.5× the next-best model."
- GDPval-AA v2 — described as state of the art.
On internal life-sciences evaluations, Anthropic puts Opus 5 10.2 percentage points above Opus 4.8 on organic chemistry and 7.7 percentage points ahead on protein tasks.
The recurring theme across CursorBench and OSWorld is cost-per-task rather than raw capability: Opus 5 is being positioned as matching or beating the far pricier Fable 5 at a fraction of the spend.
What it is actually meant to be good at
Anthropic characterises Opus 5 as a "thoughtful and proactive" model that is "much stronger at verifying its work and iterating carefully" — the behaviour that matters most when an agent is left running unattended for an extended stretch, where the failure mode is usually confident wrongness rather than an obvious crash. The company also calls it its "most aligned model to date" on the strength of a behavioural audit, and notes improved visual output.
Where it still falls short
Two limitations are stated outright. Opus 5 remains behind Mythos 5 on cybersecurity exploitation, and it is limited on long-running, autonomous research tasks in biology. Both are areas where Anthropic has been explicit about capability ceilings being partly deliberate.
What changes for developers
Two behavioural changes matter if you are moving existing code across from Opus 4.8. Extended thinking is now on by default — on Opus 4.8, omitting the thinking parameter meant the model ran without thinking at all, so a workload that never set it will now spend thinking tokens and can run into a tight output cap. And ; pairing a disabled-thinking request with or effort is rejected.