OpenAI API Cost Calculator: Which Model Is Cheapest for Your Workload?

A single-model estimator takes your token counts and multiplies. It is correct and it is not very useful, because by the time you are typing token counts into a form you have usually already decided which model to use, and that decision is the expensive one. Picking gpt-5.6-sol over gpt-5-nano on the same traffic is a difference of two orders of magnitude, which is why this OpenAI API cost calculator leads with a ranking rather than a single figure.

The reason a ranking is not just a convenience is that the order is not stable. Every OpenAI model charges more for output than for input, but the multiple varies: gpt-5.2 charges eight times its input rate for output, while gpt-4.1 and o3 charge four times. So gpt-5.2 is the cheaper of the pair on a summarisation job that reads a transcript and writes a paragraph, and the more expensive one on a drafting job that reads a brief and writes an essay. The switch happens at a specific, computable ratio.

Running that comparison across the whole line gives a number worth knowing: of the 190 possible pairs, 12 can swap places on the input to output mix alone. That is a minority, which is genuinely reassuring, and it means the 12 that do flip are worth identifying by name rather than guessing at. The OpenAI API cost calculator lists them.

Turn caching on and the picture changes again, for a reason that has nothing to do with your workload and everything to do with how OpenAI prices the discount. That is the next section.

Prompt caching bills repeated prompt prefixes at a reduced input rate. It is usually described as a 90% saving, and on the GPT-5 generation that is right. It is not right across the line, and the gap is large enough to change which model you should pick.

• 90% off, cached input at one tenth of uncached: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.4-nano, gpt-5.2, gpt-5.1, gpt-5, gpt-5-mini, gpt-5-nano
• 75% off, cached input at one quarter of uncached: gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, o3, o4-mini
• 50% off, cached input at half of uncached: gpt-4o, gpt-4o-mini, o3-mini

That accounts for 20 of the 20 models here. The remainder publish no cached-input rate at all.

The cleanest way to see why this matters: gpt-5.4 and gpt-4o both charge $2.50 per 1M uncached input tokens. Identical. At a 90% cache hit rate, gpt-5.4 has an effective input rate of $0.475 and gpt-4o has $1.375, nearly three times as much. Two models that look equally priced on the headline number are not equally priced for anyone reusing a prompt prefix.

This is also why raising the cached share in the calculator reshuffles the ranking instead of shifting every row down together. Caching cuts input cost and leaves output cost untouched, by an amount that differs per model, so it pulls the models apart. At a zero cached share, 12 pairs can flip on the token mix. At 90%, 30 can. Any OpenAI API pricing calculator that models caching as one flat percentage will get this ordering wrong.

The Batch API trades latency for price: you submit a job file, results land within 24 hours. On input and output tokens the discount really is a clean halving, and it holds for every model in this calculator. Two details are not clean, and both of them are the kind of thing a blanket percentage quietly gets wrong, so this OpenAI API cost calculator stores the batch card separately rather than deriving it.

First, gpt-5.4 publishes a batch cached-input rate of $0.13 per 1M against a standard $0.25. Half of $0.25 is $0.125, so the published cell is a 48% cut rather than exactly half, rounded to the cent. Small in isolation, and a good signal that the batch table is a separate rate card rather than a formula applied to the standard one.

Second, and much larger: 8 models publish no batch cached-input rate at all. The cell is empty for gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-4o, gpt-4o-mini, o3, o4-mini, o3-mini. There is a real difference between a rate of zero, a rate equal to the standard one, and a rate that has not been published, and only the third is true here. When you set a cached share above zero on the batch tier, this calculator moves those models out of the ranking and says why, rather than halving their standard cached rate and presenting the output as a price.

Three models were left out of this OpenAI API cost calculator for related reasons. gpt-3.5-turbo and gpt-3.5-turbo-instruct appear in the standard table with no batch row at all. gpt-3.5-turbo-1106 does have a batch row, but its batch rates are identical to its standard rates, so it is the one model here that gets no batch discount. That is a legitimate published price rather than an error, and it is worth knowing precisely because a reader who assumes batch always halves the bill would get that model wrong.

The break-even crossover is the output share at which two models cost exactly the same. It falls out of the arithmetic rather than being fitted, so this OpenAI API cost calculator solves for it directly instead of sampling. Hold the total tokens per request fixed and let s be the fraction of them that are output. Cost per request is then proportional to (1 - s) x inputRate + s x outputRate, which is a straight line in s for each model. Two straight lines meet at one point, so setting them equal and solving gives

s* = (inputB - inputA) / ((outputA - inputA) - (outputB - inputB))

That point is a real crossover only when s* lands strictly between 0 and 1. If it falls outside, one model is cheaper at every possible mix and no amount of changing your prompt shape will flip them. gpt-4o-mini against gpt-4o is that case: gpt-4o-mini wins on input and on output, so there is nothing to cross.

One edge case is worth naming because it looks like a crossover and is not. o4-mini and o3-mini publish identical input and output rates and differ only on cached input, so their cost difference works out to (1 - s) times the gap in input rate. That reaches zero only at s of exactly 1, meaning they tie solely on an all-output workload with no input at all. Equal output rates therefore rule out an interior crossover outright, and the OpenAI API cost calculator tests for that directly rather than trying to catch it with a tolerance.

In practice the useful reading is the direction. When s* is small, say the 4% between gpt-5.2 and gpt-4.1, the crossover is nearly always behind you: almost any workload that generates real prose sits above 4% output, so gpt-4.1 is the cheaper choice and gpt-5.2 only wins on near-pure classification. When s* is large, say the 78% between gpt-5.4-mini and o4-mini, the crossover is nearly always ahead of you and gpt-5.4-mini wins for anything that is not almost entirely generation.