
How Much Does the ChatGPT API Cost?
Carlos GarciaPublished How Much Does the ChatGPT API Cost?
The question has a short answer and a useful answer, and they are not the same. The short answer is that you pay per token, that a mid-range model costs a couple of dollars per million input tokens, and that most small projects spend less than a lunch. The useful answer is that the model you pick matters less than the service tier you pick, and the tier is the part almost nobody reads.
This guide is built from OpenAI's own pricing documentation as it stood on November 22, 2026. Every figure comes from OpenAI's developer pricing page rather than a third-party tracker, and where that page is silent on something an article would want to tell you, this guide says so instead of filling the gap. That matters more than usual here, because the one thing you can be certain of with API pricing is that it moves. Model names change, promotional rates expire, tiers get renamed. The structure of the billing survives; the numbers are a snapshot.
How Much Does the ChatGPT API Cost? The Direct Answer
Billing is per token, in both directions, and OpenAI states the unit plainly: "Prices per 1M tokens."
For orientation, three rates from the Standard tier as published:
gpt-6.2-sol, a mid-range model: $2.00 per million input tokens, $10.00 per million output tokens, $0.10 per million cached input tokens.gpt-6.2-astra, the larger model in that family: $8.00 input, $40.00 output, $0.40 cached input.gpt-6-luna, the cheapest in the lineup: $0.10 input, $0.50 output, $0.01 cached input.
So a request that sends 2,000 tokens of prompt and gets back 500 tokens of answer costs about $0.009 on gpt-6.2-sol and about $0.0003 on gpt-6-luna. A thousand such requests cost nine dollars on one and thirty cents on the other. That spread, roughly twenty to one between neighboring models in the same lineup, is the first thing to absorb.
There is no subscription, no seat fee and no minimum on the figures OpenAI publishes. You are billed for what you send and what comes back.
Running AI tooling into your marketing stack and not sure it is earning its keep? Our team will audit how your site and content perform in both traditional and AI search. Get your free audit.
Free audit
Where do you stand in AI search?
A manual SEO + AI search audit of your site: how ChatGPT, Gemini, Perplexity and Claude read and cite it today, and what to fix first. Report within 48 hours, no credit card.
Get my free auditOr have it all done: AI SEO services, $999 once
"The ChatGPT API" Is Not Quite What OpenAI Calls It
The product most people mean by "the ChatGPT API" is OpenAI's developer API, billed per token and accessed with an API key. The thing called ChatGPT is the consumer and business chat product, billed per user per month. They are separate purchases with separate billing, and paying for one gives you nothing on the other.
OpenAI's own pricing page does not spell that distinction out anywhere, and neither does the API platform page, which offers only "Simple pay-as-you-go pricing" and no figures at all. Across both pages there is no mention of any ChatGPT subscription plan or its price, and no sentence explaining that the two products are billed differently.
What the pricing page does contain is a row that makes the confusion worse rather than better. In its specialized-model table there is an entry labeled "ChatGPT" with the model identifier chat-latest, priced at $5.00 input, $30.00 output and $0.50 cached input. That is an API model, offered so developers can call whatever model currently sits behind the chat product. It is not the price of a ChatGPT subscription, and nothing on the page says so. If you are budgeting for a team that will use both, budget twice.
How Per-Token Billing Actually Works
Input, output and cached input
Every request has at least two priced components. Input is everything you send: the system prompt, the conversation history, the documents you pasted in. Output is everything the model generates. They are priced separately and output is always the more expensive of the two.
Cached input is the third rate and the one that changes architecture decisions. When a prefix of your request has been seen recently, OpenAI bills those tokens at the cached rate instead: on gpt-6.2-sol that is $0.10 against $2.00, and on gpt-6.2-astra it is $0.40 against $8.00. OpenAI does not state a discount percentage for caching in words anywhere on the page, so the relationship is something you read off the table rather than something the vendor commits to in a sentence.
The practical consequence is worth stating plainly: if your application sends the same long system prompt or the same reference document on every call, putting that content at the start of the request and keeping it byte-identical is the highest-leverage cost optimization available to you.
Why output costs several times input
On every model in the lineup, output is priced at roughly four to six times input. gpt-6.2-sol is $2.00 against $10.00, gpt-6-luna is $0.10 against $0.50, and gpt-5.5-pro is $30.00 against $180.00. That asymmetry should shape how you prompt: asking for a concise answer is a pricing decision as well as a quality preference, and a feature that returns long generated text will cost more than its input volume suggests.
Context length changes the rate, not just the bill
One detail that catches people building document-heavy applications. The Standard table splits on context length, with OpenAI's own footnote reading: "Short context: ≤272K input tokens. Long context: >272K input tokens." Crossing that threshold moves you to a different rate. On gpt-6.2-sol, long-context requests are priced at $4.00 input and $15.00 output against the short-context $2.00 and $10.00.
OpenAI does not define a token or explain how per-token cost is calculated or rounded. The working approximation is that a token is about four characters of English, so a thousand tokens is roughly 750 words — fine for a sketch, but measure your own traffic before you budget against it, because code, structured output and non-English text all tokenize differently.
How Do You Work Out What a Feature Will Cost?
Step 1: Count the tokens in one representative call. Not your best case and not your worst, but the one that happens most. Count input and output separately.
Step 2: Work out which input tokens repeat. A fixed system prompt, a style guide, a product catalog: anything identical across calls is a candidate for the cached rate.
Step 3: Price one call, then multiply by volume and double it. Multiply each component by its rate and divide by a million, then double the total. Retries, failed parses, evaluation runs and your own testing are all billable, and none of them appear in a forecast built from the happy path.
Step 4: Check the service tier before you commit. As the next section shows, this is the step that can halve your bill or triple it, and it is independent of the model you chose.
Want to know whether your content is showing up in AI answers at all? We audit visibility across search and AI assistants and show you where you are missing. Claim your free audit.
What Do the Current Models Cost?
All figures below are the Standard tier, short context, per million tokens, as published on OpenAI's developer pricing page. The order is OpenAI's lineup, not a recommendation.
The larger and mid-range models
gpt-6-astra: $10.00 input, $50.00 output, $1.00 cached input.gpt-6.2-astra: $8.00 input, $40.00 output, $0.40 cached input.gpt-5.5-pro: $30.00 input, $180.00 output, with no cached-input rate listed.gpt-5.5: $5.00 input, $30.00 output, $0.50 cached input.gpt-5.6-sol: $4.00 input, $20.00 output, $0.40 cached input.gpt-6.2-sol: $2.00 input, $10.00 output, $0.10 cached input.gpt-6.1-sol: $2.00 input, $10.00 output, $0.05 cached input.gpt-6.1-terra: $1.00 input, $5.00 output, $0.05 cached input.gpt-5.4: $2.50 input, $15.00 output, $0.25 cached input.
The small and cheap models
gpt-6-luna: $0.10 input, $0.50 output, $0.01 cached input.gpt-5.6-luna: $0.20 input, $1.20 output, $0.02 cached input.gpt-5.4-nano: $0.20 input, $1.25 output, $0.02 cached input.gpt-5-nano: $0.05 input, $0.40 output, $0.005 cached input.gpt-4o-mini: $0.15 input, $0.60 output, $0.075 cached input.
A naming note, because it affects how you read that list. OpenAI labels gpt-5.4-nano explicitly as a nano model but applies no size word to gpt-6-luna or gpt-5.6-luna, even though their prices sit in the same bracket. The convention changed between generations and the pricing page does not explain the new one, so judge a model by its rate rather than by whether its name contains a familiar size word.
One dated item to flag, because it expires within days of this guide being written: OpenAI's page states that "GPT-5.6 Sol's promotional pricing is available until November 30, 2026." If you are reading this after that date, treat the gpt-5.6-sol figure above as historical and check the live page.
The Service Tier Changes the Price More Than the Model Does
This is the part of OpenAI's pricing page that rewards reading. The same model, called the same way, is priced differently depending on the processing tier, and the spread across tiers is wider than the spread between adjacent models.
Batch, at half price
For work that does not need an answer now, OpenAI publishes a separate Batch table. gpt-6.2-sol on Batch is $1.00 input, $5.00 output and $0.05 cached input against $2.00, $10.00 and $0.10 on Standard, and gpt-6.2-astra drops from $8.00 to $4.00 on input. Across the rows where both tiers are listed the Batch figures are exactly half the Standard figures, though OpenAI never states that as a percentage in words, so treat the halving as an observation about the current table rather than a published guarantee. Anything on a schedule belongs here: overnight classification, bulk summarization, enrichment of a backlog, evaluation runs.
Fast, which used to be called Priority processing
OpenAI's page carries a one-line history: "Priority processing was renamed Fast mode on July 30, 2026." If you have older notes or an internal document referring to Priority processing, that is the same thing under its previous name. Fast costs more, not less — gpt-6.2-sol on Fast is $4.00 input and $20.00 output, double the Standard rate. You are buying latency.
Ultrafast, and the price rise already on the calendar
Above Fast there is Ultrafast, and OpenAI has published its future price: "Ultrafast costs 3 times the Standard rate through February 28, 2027, and 6 times the Standard rate starting March 1, 2027 (Pacific time)," for GPT-6.2 Astra and GPT-6.2 Sol.
That is unusual and worth planning around. A tier you can afford today doubles in price on a stated date. gpt-6.1-sol on Ultrafast is listed at $12.00 input and $60.00 output. If any part of your product depends on Ultrafast latency, the March 2027 date belongs in your forecast now.
Background requests, where the discount is stated in words
Unlike Batch and caching, the background discount is given as a percentage: "model token rates for eligible background requests are 30% below Standard for the 2-hour completion window and 40% below Standard for the 8-hour completion window." OpenAI limits that to a named set of models — GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna — so if you are on a model outside that list, the completion-window discount does not apply to you.
Flex, which is on its way out, and two uplifts that add 10%
Flex still has a table and a notice attached: "Flex remains available on the models listed here as we wind down the beta. We won't be adding Flex to new models." Read that as a tier not to build on; it works today on the models that have it and it will not follow you to your next upgrade.
Two circumstances raise the rate rather than lowering it. OpenAI states that "Regional processing (data residency) endpoints are charged a 10% uplift for models released on or after March 5, 2026," and applies the same 10% uplift to FedRAMP endpoints. If you have a data-residency requirement, your effective rate is not the one in the headline table, and for a regulated buyer comparing providers that 10% belongs in the comparison.
What Else Gets Billed Beyond Text Tokens?
Text is the cheap part of a multimodal application. Images are billed per token rather than per image: gpt-image-2 is listed at $8.00 per million image input tokens, $30.00 output and $2.00 cached, with text input at $5.00. OpenAI publishes no per-image price anywhere on the pricing page, pointing instead to a calculator in its image generation guide, so you cannot read a cost per picture off the pricing page at all.
Audio splits three ways. The gpt-audio model is $32.00 per million audio input tokens and $64.00 output. Text-to-speech is priced per character rather than per token — tts-1 at "$15.00 / 1M characters" and tts-1-hd at "$30.00 / 1M characters" — and live voice is priced per minute, with gpt-live-1.1 at "$0.05 / minute" active and "$0.02 / minute" standby, OpenAI noting that "GPT-Live bills voice sessions per second." Transcription is per minute of audio and cheap: gpt-transcribe at $0.0045 a minute, gpt-4o-transcribe and Whisper at $0.006, gpt-4o-mini-transcribe at $0.003. An hour of audio transcribed costs well under a dollar. Embeddings are per token and cheaper than any chat model, with text-embedding-3-small at $0.02 per million and text-embedding-3-large at $0.13.
Built-in tools are where surprise bills come from, because they are priced per call. Web search is "$10.00 / 1k calls" plus search content tokens billed at model rates, and the older web search preview on other models is "$25.00 / 1k calls" with search content tokens free. File search costs "$2.50 / 1k calls" on the Responses API plus storage at "$0.10 / GB per day" with 1 GB free. Code-interpreter containers are billed by size and session, from $0.03 for 1 GB to $1.92 for 64 GB per twenty-minute session, with OpenAI noting that "Eligible container sessions will be billed by the minute, with a 5-minute minimum per session." And the general rule: "Tokens used for built-in tools are billed at the chosen model's per-token rates," so a tool call costs its per-call fee and the tokens it produces.
Two free items are worth knowing: the omni-moderation-latest moderation endpoint is listed as "Free", and the first gigabyte of file search storage costs nothing. One restriction is worth knowing too — OpenAI states that fine-tuning is "closed to new users," so if you were planning to fine-tune, check your account's eligibility before you design around it.
Not sure which AI investments are actually moving your traffic? A free audit from our team maps what is working, what is not, and what to fix first. Request yours here.
What OpenAI's Pricing Page Does Not Say
Four absences stand out, and each one shifts work onto you.
There is no date stamp. Neither the developer pricing page nor the API platform page carries a publication or "last updated" date anywhere in its text. For a page whose entire content is prices on a product that reprices regularly, that is the most consequential gap on it: you cannot tell how current the figure you are reading is, which is why this guide is explicit about the date it was read.
There is no "subject to change" clause either. No sentence on either page reserves the right to change prices or warns that the figures are indicative. Prices clearly do change — the page itself records a tier being renamed in July 2026, promotional rates expiring in November and December 2026, and a scheduled Ultrafast increase in March 2027 — but it makes no general statement about it.
Currency is never stated. Every figure carries a dollar sign and the letters USD do not appear, which leaves both the currency and the treatment of tax unstated for a buyer outside the United States, along with any minimum spend and the usage-tier rules that govern rate limits.
And two columns in the Standard tables, labeled "30m write" and "12h write", are never explained in the page's text. They carry prices and no definition. There is also no statement of free credits or a trial for the API on either page; the only free items named anywhere are the moderation endpoint and that first gigabyte of storage.
How Does It Compare to the Alternatives?
The verdict first: for most teams, OpenAI's mid-range models are competitively priced and the decision rarely turns on the headline rate. It turns on which provider's cheapest usable model is good enough for your task, because that is where the ten-to-one savings live.
Against Google's Gemini API, the billing structure is close enough to compare directly — per million input and output tokens, cached input priced separately, context-length thresholds that change the rate. The same is true of Anthropic's Claude API. In both cases the comparison worth doing is not flagship against flagship but cheapest-acceptable against cheapest-acceptable, since a task that runs fine on a small model is where a provider choice actually shows up on an invoice.
Against running an open-weight model yourself, the arithmetic changes shape entirely. Per-token API pricing has no fixed cost and no idle cost, so self-hosting only wins at sustained high volume, and the break-even sits further out than most teams expect once engineering time is counted.
And against the ChatGPT subscription, for a team that mostly wants a chat interface rather than an integration, the subscription is usually both cheaper and less work. The API is for building something; it is a poor way to give twenty people a chat window.
Spending on AI and search without a clear picture of the return? We will audit your visibility for free and tell you where the gaps are. Start your free audit.
Final Thoughts
The ChatGPT API is billed per token, with input, output and cached input priced separately and output costing several times input. A mid-range model such as gpt-6.2-sol sits at $2.00 and $10.00 per million tokens, the cheapest models are roughly twenty times cheaper than that, and the specialized top end runs to $180.00 per million output tokens.
The two decisions that move your bill most are not which model you pick. They are whether your repeated prompt content hits the cached rate, and which service tier you run on. Batch is half the Standard price on the current table, background completion windows are 30% and 40% below Standard on four named models, and Ultrafast is three times Standard today and six times from March 1, 2027.
Treat every figure here as a snapshot read on November 22, 2026, and treat that caution as the page's fault rather than ours: OpenAI publishes no date and no change notice on its pricing page while simultaneously documenting a tier rename, two expiring promotions and a scheduled price rise. The structure is stable and worth learning; the numbers need checking against the live page before you commit a budget to them.
If you are weighing providers rather than just budgeting for one, our guide to how Google's Gemini API pricing works walks the same ground for the closest comparable product, and reading the two together is the fastest way to see which cheap model fits your task.
Want this done for you?
SEO Stuff gets businesses cited by ChatGPT, Gemini, Perplexity and Claude, and ranking in Google. Start with the free audit, a call, or the package.
See where you stand
A free, manual SEO + AI search audit of your site, with a report within 48 hours. No credit card, no sales call.
Get my free auditTalk it through
A free 15-minute call with the founder about whether AI search is a meaningful opportunity for your business.
Free strategy callHave it done
The Done-For-You Package: audit, 10 pages of content, 3 DR50+ placements, dashboard. $999 once, delivered in 21 business days.
See the package

