Anthropic's Opus 5 Matches Its Own Flagship on Coding at Half the Cost per Task

Anthropic's Claude Opus 5, released July 24, is the clearest recent evidence that capability per unit of inference compute is still improving. Anthropic says Opus 5 outperforms its prior models on Frontier-Bench, a software-engineering evaluation, and more than doubles the score of Opus 4.8 at a lower cost per task, while pricing stays at the Opus tier of roughly $5 per million input tokens and $25 per million output tokens.

The efficiency figures are what stand out. On the company's internal trading benchmark, Opus 5 reaches its result using about a seventh of the reasoning tokens and under half the latency of Opus 4.8. On CursorBench 3.2 at maximum effort, Anthropic reports that it comes within half a percentage point of Fable 5, its top model, at half the cost. Anthropic also claims Opus 5 scored roughly three times the next-best model on ARC-AGI 3, a test of novel problems, a figure that has not yet been independently reproduced.

These are vendor-reported results on a mix of public and internal evaluations, so the caution that applies to Alibaba applies here too. The difference is that Anthropic named the benchmarks and the token and latency differences, which are the numbers that matter to anyone paying for long-running agent runs, where token count rather than headline accuracy determines the cost.

Opus 5 is Anthropic's fourth model release in roughly two months, a pace that reflects how central coding and agentic work have become to the competition.