OpenAI charges for the API per million tokens, with separate prices for the input you send, the output the model writes and cached input it has already seen. As of Nov 2026 the flagship table lists gpt-6-luna at $0.10 input and $0.50 output per million tokens and gpt-6-astra at $10.00 input and $50.00 output, so the model you choose moves your bill more than anything else.
A token is a small chunk of text, and every request is billed on the tokens going in and the tokens coming out. If you're new to the API itself, my guide to what the OpenAI API is covers the basics. This page is about reading OpenAI's pricing page without getting lost in it.
How OpenAI charges for the API
Three things surprise people. First, the way you call the API doesn't change the price: OpenAI says the Responses API, Chat Completions API, Realtime API, Batch API and Assistants API aren't priced separately, and tokens are billed at the chosen model's input and output rates. Second, tokens used by built-in tools are billed at the model's rates too. Third, if you use OpenAI models through Amazon Bedrock or Microsoft Azure, you're billed through those services instead.
API pricing is also a different thing from the ChatGPT app. Using GPT-6 inside ChatGPT is covered in my guide to GPT-6 in the app, and the credits you buy for a ChatGPT plan are explained in my page on ChatGPT credits. Neither of those is the API bill.
Input, output and cached input: the numbers on every row
OpenAI's flagship table has a lot of price columns per model. They split into two halves, short context and long context, and each half has the same five numbers:
- Input: what you pay for the prompt you send.
- Output: what you pay for the text the model writes back. On every flagship row it's the bigger number.
- Cached input: a much cheaper rate for input the model has already seen, such as a long set of instructions you send every time.
- 30m write and 12h write: the cost of storing input in the cache for 30 minutes or 12 hours.
- Long context: higher prices for all of the above when a request is very long.
The cached input column is where savings hide. On gpt-6-astra, cached input is $1.00 per million tokens against $10.00 for fresh input. If your workflow sends the same long brief or product catalogue with every request, it's worth asking your developer about caching it.
Five flagship models and their prices
These rows are copied from OpenAI's pricing page as of Nov 2026, in prices per 1M tokens. The page is long and changes often, so check it before you budget.
| Model | Input | Output | Cached input | Long context input / output |
|---|---|---|---|---|
| gpt-6-astra | $10.00 | $50.00 | $1.00 | $20.00 / $75.00 |
| gpt-5.6-sol | $4.00 | $20.00 | $0.40 | $8.00 / $30.00 |
| gpt-6.1-sol | $2.00 | $10.00 | $0.10 | $4.00 / $15.00 |
| gpt-6.1-terra | $1.00 | $5.00 | $0.05 | $2.00 / $7.50 |
| gpt-6-luna | $0.10 | $0.50 | $0.01 | $0.20 / $0.75 |
One note sits under that table: GPT-5.6 Sol's promotional pricing is available at least through 21 November 2026. A promotional price can change, so don't build a long-term budget on that row without checking again.
Extras that change the bill
- Data residency: regional processing endpoints carry a ten per cent uplift for models released on or after 5 March 2026, and from 1 Nov 2026 FedRAMP endpoints do too.
- Fast mode: Priority processing was renamed Fast mode on 30 July 2026. It costs more. In the specialised table, gpt-5.3-codex is $1.75 input and $14.00 output on Standard and $3.50 input and $28.00 output on Fast.
- Voice: GPT-Live 1 voice sessions are $0.05 per minute, billed per second without rounding up to a whole minute. Backend model and tool usage is charged separately.
- Audio versus text: on gpt-realtime-2.1, text is $4.00 input and $24.00 output, while audio is $32.00 input and $64.00 output. If text will do the job, use text.
- Tools: web search is $10.00 per 1k calls plus search content tokens at model rates, and file search storage is $0.10 per GB per day after 1 GB free.
Models a small business can ignore
The pricing page also lists Cyber models (the Daybreak line) and Life sciences models (the Rosalind line). Access to the Rosalind models is limited to approved research through a trusted-access programme. They're not aimed at a marketing team or a small business, so skip those tables. OpenAI is also winding down its fine-tuning platform, which is no longer open to new users, so ignore the fine-tuning prices unless you already use it.
How I'd estimate a small business bill
I won't give you a cost per email or per blog post, because it depends entirely on how long your prompts and answers are. Here's the method I'd use instead:
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
- Pick one workflow, such as drafting replies to customer questions.
- Run it for a normal week on one of the cheaper models in the table above.
- Read the token totals from your own account's usage figures, split into input and output.
- Multiply each by the price per million tokens for that model, then repeat for a bigger model if the cheap one's answers aren't good enough.
- Set a hard spend limit before you go live. OpenAI says spend alerts send notifications without stopping API traffic, while hard spend limits can stop requests when tracked spend reaches the amount you set.
My verdict: I'd start with the cheapest model that passes your own test, and only move up when it fails. Skip gpt-6-astra until a cheaper model has let you down on the actual task. And if the numbers still look wrong for your use, my list of OpenAI API alternatives is worth a read.
If you want to forecast what an AI-powered workflow will cost before you build it, that's the kind of AI enablement I help with. I run AI training and workshops for teams, and help businesses with AI enablement and workflow automation, so the busywork runs itself and people get time back. Most of that work is with SaaS marketing teams and small businesses. Here's how we can work together.
Frequently asked questions
How does OpenAI price the API?
Per million tokens, with separate prices for input, output and cached input, and higher prices for long context. The rate depends on the model you choose.
What is the cheapest OpenAI model in the flagship table?
Of the rows in my table, gpt-6-luna is the cheapest, at $0.10 input and $0.50 output per million tokens as of Nov 2026.
What is cached input on OpenAI's pricing page?
It's a lower rate for input the model has already seen. On gpt-6-astra, cached input is $1.00 per million tokens against $10.00 for fresh input, as of Nov 2026.
Is Fast mode more expensive?
Yes. Priority processing was renamed Fast mode on 30 July 2026, and the Fast prices on the pricing page are higher than Standard.
Does the Batch API have its own price?
OpenAI says the Responses, Chat Completions, Realtime, Batch and Assistants APIs aren't priced separately. Tokens are billed at the chosen model's input and output rates.
