Cost per success
What does a successful agent task cost?
A run that fails is still billed. The price of a finished task is the price of a run divided by how often a run finishes, so a cheap model that fails often can cost more than an expensive one that does not. Enter your own success rate and see where each model lands.
No signup · Nothing leaves your browser · Success rates are yours, not ours
Price a finished task
Pick the model you use now and your measured success rate. The table shows the lowest success rate at which every other model would cost less per success. The URL keeps your settings.
Your model
Your measured rate. 80% is a placeholder.
Both are placeholders. Use your own figures.
One run
Standard rates. The batch and cache-write options on the calculator are not carried over.
Cost per success
…···
Per completed task
···
Per run
···
Success rate
···
Added by failures
Success rate a model needs to beat yours
| Model | Per run | Needs a success rate above | Cost per success at your rate |
|---|
This is a floor to test against your own measurements, not a claim about how well any model performs. "Cannot" means the model costs more per run than your cost per success, so no success rate up to 100% gets it there. A model whose last request is larger than its context window is greyed out. Long-context price tiers are not modelled here; the calculator flags them.
Your model at other success rates
| Success rate | Per completed task | Against a perfect run |
|---|
What this cannot tell you
How often any model finishes your task. The tool holds no success-rate data, and published benchmarks measure other tasks.
- It treats a run as pass or fail. A run that is partly right and needs a person to fix it is closest to the second option above.
- Retries inside a run are priced by the retry rate. The success rate is a separate, whole-run outcome.
- Time spent checking output, and upkeep, are not priced.
Questions
What is cost per successful task?
A failed run is still billed, so the cost of a finished task is higher than the cost of a run. If failed runs are simply run again, cost per success is the cost of one run divided by the success rate. If a person finishes the failed ones, it is the cost of a run plus the failure rate times the person's cost per task. On the default workload a run costs $0.351 on Claude Sonnet 5.5, so at an 80% success rate each completed task costs $0.439, which is 25% more than the run price.
How low can a cheaper model's success rate go before it costs more per success?
Take the cheaper model's cost per run divided by your current model's, and multiply by your current success rate. On the default workload GPT-6 Luna costs $0.018 a run against $0.351 for Claude Sonnet 5.5, a ratio of 0.05, so against an 80% success rate it only needs to complete 4.0% of runs to cost less per success. That is a floor to test with your own measurements, not a statement about how well any model performs. A model that costs more per run, such as Claude Fable 5.1 at five times the price, would need a success rate of 400%, so it cannot win on cost per success at any rate.
Where do the success rates come from?
You enter them. The tool holds no success-rate data for any model, and published benchmarks measure different tasks from yours. Run a sample of your real tasks, count the ones that pass your own acceptance check, and enter that share. Cost per success is only as good as that number.