Tests find Claude Sonnet 5.5 as good as Opus 5.5 at half the cost

A test in October 2026 gave 3 models from Anthropic the same 11 code tasks and 1 repair task in Claude Code. The models are Opus 5.5, Sonnet 5.5 and Haiku 5.5. All 3 got all tasks right. Sonnet cost between half and 2 thirds of the cost of Opus. The testers recommend Sonnet 5.5 as the default for most teams.

How it was tested

The testers wrote 11 Python tasks with test files that the models cannot see. 3 of the tasks were hard. They also gave each model 3 bugs to repair in a small invoice library. Each model used its default.

The numbers

Sonnet got 92.4% on a science test, and Opus got 90.4%. On an agent test with customer questions, Sonnet got 78.7% and Opus got 75.5%. On that test Sonnet cost $0.16 for each task, and Opus cost $0.34. Haiku got 76.0% at $0.011 for each task.

What is not known

On the Artificial Analysis index, a score from 10 tests, Opus has 1.6 more points. The tasks were short. They cannot show the result of Opus on long tasks that are not clear, where an error has a high cost.

Insight: The default model can send your cost to the dearest tier

Opus 5.5 is the default model in Claude Code on most plans. In the week of October 12, it had about 2.5 times the token volume of Sonnet on OpenRouter, a model marketplace. Opus also has the top score, 52%, on the FrontierCode 1.1 test. Maybe a top score on hard tests and a default set by the maker move cost to Opus, not the work.