Compare models

Type part of a model name and pick from the list. The lowest price in each row is highlighted. The link updates as you choose, so you can share it.

Reliability: cost per successful task

Assumptions: attempts are independent and cost the same. A failed final attempt is still paid. Latency is not modelled.

Monthly costs use the same scenarios as the calculator (30 days). Edit this workload in the calculator →

Cheaper per attempt is not always cheaper per finished task. If a task is retried until it succeeds, the expected cost per success is the cost per attempt divided by the success rate per attempt (C ÷ p). Capping the attempts lowers the bill but also lowers the share of tasks that finish. Enter success rates from your own evaluation in the Reliability section; the tool never fills them in.

FAQ

Which model is the best value?

It depends on the task. Compare the monthly cost for a scenario close to yours, then test the two or three cheapest models on your own prompts before deciding.

Why do some cells show —?

The provider does not publish that price (for example no batch API or no long-context tier) or it is missing from our data source.

Why can a cheaper model cost more per finished task?

Retries are paid too. Cost per successful task = cost per attempt ÷ success rate per attempt. Hypothetical example, not real models: model A costs $0.010 per attempt and succeeds 95% of the time, so $0.010 ÷ 0.95 = $0.0105 per success. Model B costs $0.006 per attempt but succeeds 50% of the time, so $0.006 ÷ 0.50 = $0.012 per success. B is 40% cheaper per attempt yet about 14% more expensive per finished task. With attempts capped at 3, B also finishes only 87.5% of tasks (1 − 0.5³) versus 99.99% for A.