A three-task API comparison published on September 26, 2026, found that Claude Opus 5.5 and Claude Opus 5 both solved a logic-grid puzzle and a stone game, but neither answered a constrained-ordering problem. Opus 5.5’s aggregate output rate was about 11% higher in that test.
How the comparison was run
The comparison used identical prompts sent through the Anthropic API, with adaptive thinking set to its default effort level. It ran each problem once on each model. The prompts covered a logic grid built around 22 clues, a constrained job-ordering puzzle, and a stone-removal game.
Results across the three tasks
| Task | Claude Opus 5.5 | Claude Opus 5 |
| Logic grid | 28/28 | 28/28 |
| Constrained ordering | No answer at either tested output limit | No answer at either tested output limit |
| Stone game | 3/3 questions correct | 3/3 questions correct |
The ordering puzzle asked for valid arrangements of 6, 8 and 10 jobs under constraints. Neither model answered it at the tested 48,000- and 128,000-token output limits. In the stone game, both answered all three questions correctly.
Measured speed and Anthropic’s claim
Across the reported API calls, Claude Opus 5.5 generated 103.4 output tokens per second, compared with 93.1 for Claude Opus 5—about an 11% difference in this test. Anthropic said in its September 22 announcement that Opus 5.5 generates output more than 30% faster than Opus 5 on typical workloads. The two figures describe different test scopes.
The comparison covered three bespoke prompts, with one run per problem and model. Its results describe those tasks and settings; they do not establish a general ranking of the models’ reasoning ability.
