Type your token counts, pick a model, and watch the arithmetic unfold. Compare every current frontier model side-by-side, scale to production volumes, and see where caching and batching bend the curve.
| Model | Provider | In / Out per 1M | Your cost | Relative | vs. selected |
|---|
Cheapest result highlighted. Bar widths scale to the most expensive option on screen. Modifiers (caching, batch) apply to every row.
A token is the atomic unit a model reads or writes — usually a short run of characters, often a whole short word, sometimes a fragment. As a rough ratio, four English characters or about three-quarters of a word equals one token.
Generating text is the expensive part. Reading your prompt is largely parallel work; producing each output token requires another forward pass through the model. Most providers price output three to five times higher than input.
If your requests share a long system prompt or document, providers can cache those tokens so repeated requests only pay a tiny fraction of the usual input rate — typically a 90% discount on the cached portion. The single biggest lever for agent-style workloads.
Anything that doesn't need to answer a human within seconds — evaluations, bulk data processing, overnight content pipelines. Batch jobs return within 24 hours at half price on both input and output tokens, every model.
Route by difficulty. Send classification and extraction to the smallest model, give the mid-tier the bulk of your traffic, and reserve the flagship for the requests where reasoning quality genuinely matters. A blended bill is almost always cheaper than a uniform one.
Rates are sourced from each provider's published pricing as of May 2026. Currency conversions are approximate. Confirm against your provider's current rate card before forecasting any material budget.