Cheapest LLM API: How to Actually Compare Prices

There is no cheapest LLM API. There is a cheapest one for the work you are doing, and it is a different model depending on what that work looks like. We can show this rather than assert it. Priced on published rates for 42 models on August 10, 2026, the cheapest model for a summarizing workload is GPT 5 Nano, the cheapest for a writing workload is Ministral 3 3B 2512, and the cheapest for balanced chat is Mistral Small 3.2 2506. Three answers, one question, nothing changed but the shape of the job.

The same model, two jobs, a different answer

Input and output are metered separately and priced differently, so a ranking built on one is wrong for the other. Take GPT 5 Nano. On a summarizing job that reads 10,000 tokens and writes 200, it is the 1st cheapest of the 42 models we price. On a writing job that reads 200 and writes 2,000, the same model on the same rate card is only the 8th. Nobody repriced anything. The work changed direction.

The spread is not a rounding error either. On that summarizing job the most expensive model we price costs 190 times the cheapest. On the writing job the gap widens to 464 times. Picking from a sticker-price list is choosing somewhere inside a range that wide, with your eyes closed.

Work out which direction your tokens flow

You can settle this in a minute without any tooling. Take one typical request. Count roughly what goes in: your system instructions, any documents or examples you attach, the conversation so far, and the user's actual message. Then count roughly what comes back. If the first number is several times the second, you are running an input-heavy job and input rates decide your bill: summarizing, classifying, extracting fields, answering from documents. If the second is larger, you are output-heavy and output rates decide it: drafting, writing, generating code. Balanced chat sits between the two and is the only shape where the sticker headline is close to the truth.

One warning about the second number. Reasoning-capable models bill their thinking as output, and that thinking is not visible in the reply you measured. On those models an output-heavy job can cost several times what the reply length suggests, which is the single most common reason a real bill overshoots an estimate.

Caching changes the ranking again

If any part of your prompt repeats between calls, and for most production systems the system prompt, the tool definitions and the shared documents all do, then the rate that matters for that part is the cache-read rate, not the input rate. Of the 42 models we price, 28 publish a cache-read rate and 14 do not. Among those that do, the median discount is 90 percent off standard input.

Two consequences follow, and both invert the usual advice. A model with a higher sticker price and a deep cache discount can beat a nominally cheaper model on any prefix-heavy workload. And a model that publishes no cache-read rate is more expensive than its sticker suggests for that same workload, because the repeated part never gets cheaper no matter how often you send it. If you are building something with a large fixed prefix, filter by cache support before you compare headline rates at all.

The cheapest model that fails is not cheap

We publish no quality score, and we are not going to. Every figure on this site is a published rate or arithmetic on one, and a quality ranking is neither. So the last part of this decision is yours, and it is the part no price list can make for you: how wrong can an answer be before it costs you more than the model saved?

The practical test is cheap to run. Take fifty real inputs, run them through your shortlist, and read the outputs yourself. A budget model that handles forty-eight of them is almost always the right answer at these prices. A budget model that handles thirty is the most expensive option on the page once you count the retries, the corrections, and the person checking. Cost is only decidable against a quality floor, and only you know yours.

Find yours

Three routes, depending on what you already know. If you know your token counts, the rates page ranks every model we price by computed cost for four workload shapes, so you can read your answer off the shape that matches. If you have a real prompt, paste it into the cost calculator and the top row is your answer for that exact text. If you are sizing something that does not exist yet, the project estimator asks four plain questions and prices the whole thing, and it will stack several workloads into one project total.

Whichever you use, re-check it after any month in which the providers move. Ours is a live archive precisely because this answer does not hold still: the models that won the three shapes above are the winners on August 10, 2026, and the point of this guide is the method, not the names.