LLM API pricing reference

How to read this table

Every figure is US dollars per million tokens at the on-demand rate, which is the price you pay with no commitment, no volume agreement and no batch discount. It is the rate almost everyone actually pays, and the one providers quote. A token is roughly four characters of English prose, so a million tokens is somewhere near 750,000 words, or about nine copies of a full-length novel.

Input is what you send: your system prompt, your tools and schemas, any documents you supply, and the whole conversation so far, all of it re-sent on every call. Output is only what the model writes back. Cached input is what a repeated prefix costs on the second and later calls, and it is the column most people skim past and should not, because it is where the largest discounts on this page are: 32 of the 40 models publish a cache-read rate, and the median one reads at 90% below its own standard input rate.

The cheapest rate is not the cheapest bill

Sorting this table by input rate answers a question almost nobody has. Input rates across these 40 models span $0.05 to $10 per million, a factor of about 200, but output is priced separately and usually much higher: the median model here charges 5.0 times more for output than for input, and at the extreme the multiple reaches 8. So the model that wins on the input column can lose badly on a workload that writes a great deal, and the reverse is just as true.

What decides your bill is the mix. A retrieval or classification workload reads enormously and writes a sentence, so its cost is nearly all input and caching is worth more to it than any model switch. A drafting or summarizing workload writes far more than it reads, so output rate dominates and caching barely helps. Two teams paying the same provider can rank these models in opposite orders and both be right. That is why the model pages price four different workload shapes rather than publishing one ranking, and why the estimator asks about your work before it names a number.

What is in this table, and what is not

These are first-party text models from 7 providers, priced at the rate the provider that built the model charges to serve it. Resellers, aggregators and hosting marketplaces are left out, not because they are worse but because their prices move for reasons that have nothing to do with the model and everything to do with the intermediary, which would make a price archive meaningless. Fine-tuned, embedding, image generation and audio models are out of scope here too.

Every rate on this page was read from a provider's own published pricing in a dated capture and committed to the archive that night. Nothing is estimated, interpolated or projected, and when a provider publishes no rate for something, this site says so rather than filling the gap with a plausible figure. That is the whole discipline behind the archive, and it is why a cell can be empty here when other comparison sites show a number.

Rates move less often than the coverage suggests

Only 14 of the 40 models here have repriced at all since entering the archive. Providers overwhelmingly prefer to launch a cheaper new model than to cut the price of one already in service, which is why the way costs fall in practice is that a new name appears at a lower rate and you migrate to it, not that the number next to your current model gets smaller while you sleep. The practical consequence is that watching your own model's price is close to useless, and watching what arrives beside it is where the saving is. That is what the monthly index tracks.