Claude Haiku vs Sonnet vs Opus 5.5: cost per solved task on the same 27 tasks
On the same 27 coding tasks with hidden tests, Claude Sonnet 5.5 solved as many as Claude Opus 5.5 (26 of 27) for 49% less per solved task: $0.152 against $0.297. Claude Haiku 5.5 was 16 times cheaper again at $0.009 per solved task, but solved 22 and failed 4 tasks the other two passed.
The results
The open-source coding agent Aider got 27 Python exercises from its public polyglot benchmark. The agent never saw the tests. Each exercise ran once per model, with one retry using the test output, as in Aider's own benchmark. Only the model changed.
| Model | Model ID | Solved (of 27) | Solved first try | Cost per solved task | vs Opus |
|---|---|---|---|---|---|
| Claude Opus 5.5claude-opus-5-5 | claude-opus-5-5 | 26 | 26 | $0.297 | — |
| Claude Sonnet 5.5claude-sonnet-5-5 | claude-sonnet-5-5 | 26 | 25 | $0.152 | −49% |
| Claude Haiku 5.5claude-haiku-5-5 | claude-haiku-5-5 | 22 | 15 | $0.009 | −97% |
Total cost for all 27 tasks: Opus 5.5 $7.71, Sonnet 5.5 $3.95, Haiku 5.5 $0.20. Cost = tokens × list price per 1M tokens (input / output): Haiku 5.5 $0.10 / $0.50, Sonnet 5.5 $2 / $10, Opus 5.5 $4 / $20. Run on 11 November 2026 with Aider 0.86.2. All three models failed the same one exercise.
What Haiku 5.5 got wrong
For each task Haiku failed and the others passed, we re-ran the hidden tests on its final code:
- 11 / 54a small language interpreter: arithmetic and division-by-zero handling
- 8 / 9a small REST API: balances between users
- 3 / 12a graph parser: errors for malformed input
- 1 / 7a word game: the same correct letter twice counts as a miss
Failing tests out of all tests for that exercise.
The main path worked. Edge cases and error handling broke. Haiku also needed the retry far more often: 15 tasks solved on the first try, against 25 for Sonnet and 26 for Opus.
Price per token is not price per task
- Opus 5.5 vs Sonnet 5.5: twice the list price per token, 1.95 times the cost per solved task. Same tasks solved, so the extra spend bought nothing on this set.
- Sonnet 5.5 vs Haiku 5.5: 20 times the list price per token, but only 16.5 times the cost per solved task, because Haiku solved fewer tasks: each solved task also carries the cost of the failed ones.
- Cost per solved task, together with the list of failed tasks, is the number to decide on.
Which model for which job
- Haiku 5.5 is enough for high-volume work where the main path is the job and a checker catches misses: tests, schema validation, exact-match extraction, a retry loop. Don't give it edge-case-heavy logic without a check.
- Sonnet 5.5 is the default for coding agents and multi-step tool use. It matched Opus 5.5 task for task here at half the cost.
- Opus 5.5 earns its price only on the tasks where your own tests show Sonnet 5.5 failing a critical case. Route those tasks to Opus, not the whole workload.
One public benchmark; your tasks decide. That is what the pilot measures.
Method and raw data
The protocol was fixed and committed before the first paid call. Read the full report, the protocol and the raw results (CSV, one row per task and model). How we test every setup: methodology.
Moving off a retiring model? Claude Sonnet 4.5 retires on 30 November 2026; every other date is in the retirement calendar.
Find your cheapest model that still does the job
Send us 20–30 of your real tasks with expected answers. We run Haiku 5.5, Sonnet 5.5 and Opus 5.5 on them and report cost per finished task and every task each one gets wrong. The pilot is free; the report is yours.
Questions people ask
Is Sonnet 5.5 good enough instead of Opus 5.5 for coding agents?
On our 27-task Aider test, yes: Claude Sonnet 5.5 solved the same 26 tasks as Claude Opus 5.5 at $0.152 per solved task versus $0.297, 49% less. Check your own critical cases before moving everything.
How much does it cost per task to run an agent on Sonnet vs Opus?
In our test, Claude Opus 5.5 cost $0.297 per solved task and Claude Sonnet 5.5 cost $0.152, with 26 of 27 tasks solved by each. Claude Haiku 5.5 cost $0.009 per solved task with 22 solved. Cost is tokens times list price.
Claude Haiku vs Sonnet: is Haiku 5.5 good enough?
For main-path work with a checker, often yes: Haiku 5.5 was 16 times cheaper per solved task than Sonnet 5.5. It failed 4 more of 27 tasks, all on edge cases and error handling (interpreter arithmetic, REST balances, graph-parser errors, a word-game rule). Test it on your own edge cases first.
When should I use Opus 5.5 instead of Sonnet 5.5?
When your own tests show Sonnet 5.5 failing a task that matters. On our test Opus 5.5 solved nothing Sonnet 5.5 missed and cost about twice as much per solved task, so route only the tasks that need it to Opus.
Why compare cost per solved task instead of price per token?
Because retries and failures cost money too. Sonnet 5.5 has 20 times Haiku 5.5's list price per token, but only 16.5 times its cost per solved task on our test, because Haiku solved fewer tasks and each solved task carries the cost of the failed ones.
How do I find out which Claude model is right for my tasks?
Run 20–30 of your real tasks with expected answers on each model, compare cost per finished task, and list the tasks each model gets wrong. Our free pilot does exactly that and returns a one-page report within a week.
Questions or corrections: hello@agenttrustapi.com. Every public number here comes from our Aider report and its raw data.