Why output tokens usually dominate the bill
Every major provider charges input and output tokens at different rates, and output is almost always priced several multiples higher than input — which means a chatty, verbose model response costs disproportionately more than a longer prompt sent to a concise one. This is the single most common miscalculation teams make estimating AI costs: they budget based on prompt size and ignore that response length is usually the bigger lever. The table above lets you see exactly where your specific input/output mix lands across all three model families at once.
The practical cost-optimisation strategy most production systems use isn't picking one model — it's routing by task: cheap, fast models (the "Flash," "Haiku" or "Luna"-tier options here) handle high-volume, simple tasks, while an expensive flagship model is reserved for the smaller share of requests that genuinely need its extra capability. This calculator's per-model comparison is exactly the input to that kind of routing decision.
None of this includes prompt/context caching or batch processing, both of which can cut costs significantly for the right workload pattern — treat the numbers above as a ceiling to budget against, not the lowest achievable cost. See AI automation for how we typically architect model selection and cost control into a production system, and custom software for the integration layer around it.