FIELD NOTES / SEPTEMBER 2026
Sonnet 5.5 vs Opus 5.5: Cost and Test Plan
Sources checked
Choose between Sonnet 5.5 and Opus 5.5 by testing the same task with the same inputs and acceptance criteria. A lower token price does not necessarily mean a lower cost per accepted result.
Anthropic positions Sonnet for well-scoped everyday work and Opus for more complex tasks. These are vendor descriptions, not our test findings. Official announcement and pricing table.
Published standard API rates
| Per million tokens, USD | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Uncached input | $2 | $4 |
| Output | $10 | $20 |
| Cache reads | $0.20 | $0.20 |
Checked September 29, 2026. The input and output rates above are half as much for Sonnet; cache-read rates are equal. Total task cost depends on token usage and retries. Subscription prices are separate.
Estimate uncached token cost
Same token counts for both models. Excludes cache writes, tools, search fees, discounts, taxes and any extra usage. Enter billable output tokens from your provider.
A test you can reproduce
- Task: fix a keyboard-navigation issue in the same small website checkout.
- Inputs: identical files, prompt, tools, time limit and starting commit.
- Acceptance: keyboard access works, existing tests pass, and no unrelated files change.
- Repeat: run each model three times in separate clean sessions. Record every attempt, not just the best.
- Measure: completion, human correction time, elapsed time, billed tokens and cost per accepted result.
Download blank results worksheet
Record model IDs, effort settings and test date. If one model needs a retry, include it in that model's cost. Keep provider benchmark results separate from your own measurements.
Which should I try first?
For a narrow task with a clear pass/fail check, start by testing the cheaper option. For ambiguous work, compare correction effort as well as price. This is an evaluation strategy, not a claim that either model wins your workload.