Start with the billing units
Some APIs split input, cached input, output and reasoning rates. Others publish one blended rate for charged tokens. Put both into a single comparison only after defining the token mix and minimum-charge rules.
For a split tariff, calculate each token class separately. For a blended tariff, multiply the total charged tokens by the blended rate and then apply the documented minimum request charge. Do not compare a blended rate with an input-only rate.
Use one representative workload
Create a small test set that resembles your application: a short question, a coding task, an input-heavy context and an output-heavy task. Keep the prompts synthetic and record the exact model ID, token budget, cache setting and reasoning level.
Repeat the same requests against every candidate. A single short prompt is useful for a smoke test, but it cannot tell you which service is cheapest for an agent that sends long contexts and retries failed streams.
Measure the final charge
The number in a pricing table is an estimate until a request is settled. Record the terminal input, cached, output and reasoning tokens, the final charge, the minimum charge and any retry or payment fee.
For a prepaid wallet, verify that the temporary reservation is released after settlement. If terminal usage is missing, the safe behavior is reconciliation rather than guessing a token total.
Streaming changes the decision
Interactive coding tools care about time to first byte and time to first token, not only total latency. Measure P50 and P95 latency, stream completion, disconnect handling, 429 responses and 5xx responses next to cost.
A service that is slightly cheaper but repeatedly breaks streams can cost more through retries, support work and lost user time. Test cancellation explicitly and check that the request is either settled or left in a documented pending state.
Official API, aggregator or prepaid gateway?
The official API is usually the clearest choice when you need provider-native features and direct contractual support. An aggregator is useful when breadth, routing and provider choice matter. A prepaid compatibility gateway is convenient when one key and a fixed wallet are more important than a large catalog.
Terms, model catalogs, fees and limits change. Check the current documentation and commercial-use policy for the service you choose; SDK compatibility does not imply ownership or affiliation.
A reproducible checklist
Choose two or three models, define the workload, run JSON and streaming, record terminal usage, calculate the final charge, measure latency and repeat the test after at least 24 hours. Publish the check date and assumptions with any comparison.
This is the method used for the Model.sale public catalog: prices remain visible with their units, while API admission is limited to models with a recent compatibility check and a working usage policy.