Anthropic released Claude Sonnet 5.5 on 28 September, six days after Claude Opus 5.5. We have been deciding which one to use for our own work, and our view is simple. For most tasks, Sonnet 5.5 is enough.
The prices
From Anthropic's pricing page, per million tokens: Sonnet 5.5 costs $2 for input and $10 for output. Opus 5.5 costs $4 and $20. Reading from the prompt cache costs $0.20 on both, and the Batch API halves the input and output rates. So per token, Opus is exactly twice the price.
The benchmarks
Anthropic published the two models side by side. Sonnet 5.5 comes out ahead on one test, Terminal-Bench 4.0, which measures agentic coding in a terminal: 70.6% against 66.4% for Opus. Opus leads everywhere else in the table, mostly by a little. On OSWorld 2.1, a test of operating a computer, it is 81.8% against 80.1%. On CursorBench 4.0 it is 57.8% against 55.5%. On GDPval-AA v2.1, which scores professional knowledge work on a rating scale, it is 1846 against 1844, a tie in practice.
The clearest gap is FrontierCode 1.1, a harder coding test. Opus scores 54.4%. Sonnet scores 46.2% at max effort and 52.1% at the xhigh setting, one step down. On Humanity's Last Exam with tools, Opus leads 67.7% to 64.5%.
Vendor numbers deserve a second opinion, so we checked Artificial Analysis, which runs its own tests. On its Intelligence Index, Opus 5.5 at max effort scores 58 and Sonnet 5.5 scores 56. In its own Terminal-Bench run Sonnet again came out ahead, 64% to 60%. The widest gap it found was factual accuracy on its AA-Omniscience test: 66% for Opus, 54% for Sonnet. Sonnet was also faster, at 139 output tokens per second against 92.
The catch in the price
The price per token doesn't tell you the bill, because the bill also depends on how many tokens the model writes. Artificial Analysis found that at max effort Sonnet 5.5 used about 193,000 output tokens per task on its index, against 119,000 for Opus. Running the whole index cost $8,977 on Sonnet and $8,708 on Opus. At max effort, Sonnet saved nothing.
So we don't run Sonnet at max effort by default. Anthropic says Sonnet 5.5 costs up to 30% less per task than Sonnet 5, but that is a comparison with its predecessor, not with Opus. Measure tokens per task on your own work before you count on a saving.
When Opus is still worth it
Anthropic's own guidance is fair: Opus 5.5 is for complex work that needs careful judgment, and Sonnet 5.5 is strongest at well-scoped everyday tasks, bug fixes and documents. The numbers agree. We would use Opus for long jobs where one wrong turn early wastes hours, such as a change that runs across a large codebase. We would also use it where facts have to be right and nothing downstream checks them, given the 12-point gap on factual accuracy. On hard coding problems of the FrontierCode kind, the gap at max effort is about 8 points, although Sonnet at xhigh narrows it to about 2.
For drafting, summarising, routine code changes, spreadsheet and slide work, and agent steps with a clear test of success, we don't see a reason to pay double.
Our rule of thumb
Start with Sonnet 5.5 at a normal effort level. Write a test for the task, even a small one: a handful of inputs and the answers you would accept. If Sonnet passes, stay there. If it fails, raise the effort, then try Opus. Keep the test, and run it again when a new model comes out.
The test matters more than the rule. Published benchmarks tell you where to start looking, and your own task decides which model is good enough for it.