OpenAI API Token Rates, Batch Savings & the Real Bill 2026 Guide

The OpenAI API has no seat price, so the whole bill is tokens, and the single biggest lever is which model you call. Input runs $0.15 per million on GPT-5.5 Mini and $15 per million on o1, a hundredfold spread. The same 10-million-token job is a dollar-fifty on Mini or a hundred and fifty on o1. Volume is not what sets the invoice. Model selection is.

Output is the second trap, because every model charges more to write than to read. Flagship GPT-5.5 is $5 per million input and $30 per million output, and o1 climbs to $60 per million out. A chatty, long-response application spends far past what the input rate implies, so the honest way to budget is from the tokens the model generates, not the ones you send.

Two discounts pull the effective rate back down, and both are easy to miss. The Batch API takes 50 percent off every model for jobs that tolerate a delayed answer, and prompt caching takes another 50 percent off repeated input on a cache hit. Above that sits provisioned throughput for dedicated capacity, which is quote-only. The full rate card lives on the OpenAI API pricing page, and there is no monthly cap, so a runaway loop bills until you set a hard limit yourself.

There is no free plan in the usual sense, because there is no plan. The OpenAI API is pay-as-you-go, so an idle account costs $0 and you pay only while requests run. New accounts also get a small pool of trial credits to prototype against before any card is charged, and those credits expire, so use them while they last.

That structure is friendlier than a subscription for light or spiky work. You are never paying for a seat you did not use. It is harder for heavy production traffic, where the absence of a monthly cap means the bill tracks demand exactly. Build and test on the trial credits, confirm the model fits, then set a hard spend limit before you ship. If the rate card gives you pause, cheaper token rates sit a click away on the OpenAI API alternatives page.

There are no seats here to discount and no coupon codes to hunt. OpenAI has never offered a student, academic or nonprofit rate on the API, because a token meter gives a coupon nothing to bite on. What exists instead are engineering discounts, and they are larger than most people realize.

Send delay-tolerant work through the Batch API and every model rate halves. Reuse a large fixed context and prompt caching halves the repeated input on a cache hit. Stack the two and the effective rate drops well under the headline. Beyond that, provisioned throughput and committed-use terms are negotiated for organizations moving real volume, which is what the negotiation section covers.

At the published rate card there is nothing to negotiate. The prices are the prices, and the levers are engineering ones: batch, caching and model choice. Negotiation begins only when your committed volume is large enough for OpenAI's sales team to care, which is the point where provisioned throughput and custom rates come into play.

Three moves carry most of the weight. Each rests on one fact: OpenAI would rather lock in your revenue than watch you meter it call by call.

Day-to-day API spend has no timing angle. The rate card ignores the calendar, so a committed-use conversation is the only thing worth timing, and it follows the usual sales rhythm. OpenAI's enterprise reps carry quarterly quotas, so a committed-spend deal closing in the final weeks of a quarter usually prices sharper than one opened at the start.

Aim requests at the part OpenAI can actually flex. On the API the rate card is fixed for everyone below committed volume; the room opens with reserved capacity and a spend guarantee.

Fill in the details and the draft below assembles itself, pulling live rival token rates from our verified catalog. Send it to your OpenAI sales contact or the enterprise inquiry link. The sequence is what matters. Open with your monthly token volume, lay a competing rate card next to it, request committed-use pricing, and name a start date the deal can hit.

Each of these traces back to the way token billing works, and all of them can be shut before your first production call.

On a usage-priced API, leverage is arithmetic. These three model houses publish rate cards you can hold up next to OpenAI's, verified in our catalog. You do not need to migrate. You need a competing number you have actually benchmarked, so the comparison is real rather than rhetorical when you sit down with sales.