Opus 5.5: Cheaper Tokens, More Expensive Tasks
In the reported coding benchmark, performance increases. At maximum reasoning effort, however, so does the bill. What matters are the costs per successfully completed task. Published on 25 September 2026 • AI translated
A lower token price does not automatically make a coding agent cheaper. Claude Opus 5.5 demonstrates precisely this difference: in the reported test, the model achieves the highest overall score, but costs 21 percent more per benchmark task than Opus 5.
This is not a contradiction. The test was run in Claude Code with maximum reasoning effort, meaning the highest configured effort setting for reasoning. In this configuration, Opus 5.5 consumes so many additional tokens that the lower unit prices cannot offset the added costs.
The performance gains in the described test are clear: the overall score climbs to 66 points, six points ahead of Opus 5. All three sub-tests show gains. The largest leap is made by Terminal-Bench 4.0 with 8.6 percentage points, rising from 54.5 to 63.1 percent.
This sub-test evaluates, among other things, command-line operations, proper tool usage, and error recovery. That is relevant for longer coding tasks. However, it does not yet answer whether the extra performance pays off in your own workflow.
| Metric in the reported test | Opus 5 | Opus 5.5 |
| Estimated API cost per task | 10.79 US dollars | 13.04 US dollars |
| Total tokens per task | Around 11.4 million | Around 15.6 million |
| Output tokens per task | Around 137,000 | Around 333,000 |
The increase in output is particularly striking: from around 137,000 to 333,000 tokens per task. At the same time, the stated standard prices for input and output both drop by 20 percent. Cheaper units, significantly higher consumption. The bottom line remains a higher bill.
And what about the announced 40 percent savings? The cited vendor estimate applies to typical workloads under default settings, not to this benchmark run at maximum effort. Both claims can be true at the same time because they describe different conditions.
What is missing is a controlled comparison of the same tasks across multiple effort settings. The available results do not quantify how much performance remains at lower effort levels and what savings actually materialize as a result. Therefore, no blanket recommendation for production use can be derived from the peak score.
For teams, a different metric is decisive: What does a successfully completed and accepted task cost? The 13.04 US dollars represent estimated API costs per benchmark task. They are not an established price per successful production task including retries and human review.
1. Compare your own tasks
Select representative tasks from your own repositories. Test the same prompts and baseline conditions across multiple reasoning settings, rather than merely pitting the maximum setting against the previous model.
2. Track outcomes and effort
Log accepted completions, retries, correction runs, and execution time. In addition, measure input, output, and cache usage, as well as total API costs. A single successful run is not a sufficient basis for decision-making.
3. Calculate by successful completion
Divide the total API costs of the test runs by the number of accepted completions. Factor in review effort and waiting time as well. This makes it possible to assess whether higher performance justifies its added costs in your specific workflow.
Optimize for the result, not the token price
In the reported test, Opus 5.5 shows higher coding performance at higher costs per task. Whether that is a good trade-off is not decided by the leaderboard. What matters is which setting completes your tasks reliably, with acceptable waiting times and the lowest total cost per accepted result.