Claude Opus 5 keeps the price but roughly doubles the score

Anthropic released Claude Opus 5 at exactly the price of the model it replaces: $5 per million input tokens and $25 per million output tokens. The number that matters is not the price but what you get for it. On Frontier-Bench v0.1 the new model more than doubles Opus 4.8's score, and on CursorBench 3.2 it lands within 0.5 percent of Fable 5 while costing half as much. On OSWorld 2.0, a test of an agent driving a real computer, it beats Fable 5 at about a third of the cost.

The pitch is agentic work that checks itself. Anthropic says Opus 5 is better at verifying its own output and iterating instead of stopping at the first plausible answer, which is where long agent runs usually fall apart. One example in the announcement describes the model writing its own computer vision pipeline when it could not read the task requirements directly, rather than giving up. The gains extend past code into scientific work, with specific improvements called out in chemistry and protein analysis, plus stronger image and diagram generation. A Fast mode runs about 2.5 times quicker at twice the base rate, so you can trade money for latency when a job is waiting on the model.

Why it matters

If you run coding agents or long automation loops, the interesting move is that the cost per finished task dropped while the price per token stayed flat. Before you commit to a more expensive frontier model, it is worth re-benchmarking Opus 5 on your own workload, because half the cost at near-parity changes the math on which model you default to.