Reasoning usage is easy to miss
Some reasoning models spend tokens on internal reasoning as well as the visible answer. Provider usage accounting and output limits differ. Inspect the output or reasoning fields in actual responses, then price the billable usage according to the provider's current documentation.
Build a representative test set
Collect 30–100 real tasks with clear success criteria. Include ordinary cases and costly failures. Run each candidate model with the same instructions and relevant settings. For each result, log total billable cost, correctness, latency and whether a retry or human correction was required.
Calculate cost per successful task
If model A costs $0.01 per attempt and succeeds 70% of the time, its simple cost per success is about $0.0143. If model B costs $0.03 and succeeds 95%, its cost per success is about $0.0316. B costs more per successful result, but it may still win if errors create meaningful review or business costs. This is an illustrative calculation, not a benchmark claim.
Route selectively
Many teams can use a cheaper model first and escalate only uncertain or high-stakes requests. Define an objective escalation rule and measure the added requests. A model choice is strongest when it is based on your tasks, quality threshold and observed token mix rather than a public leaderboard rank.
Use the tool for the price side
Tokens to USD converts measured input and output to a base text cost. Add reasoning, cache, tool or tier charges if the provider bills them separately. Then compare the complete cost with your quality results.