OpenAI API Pricing (November 2026)

Prompt caching is on by default for OpenAI models. When a request starts with the same prefix as an earlier one, the reused tokens bill at the cached-input rate instead of the full input rate.

On GPT-5.6 and later models, a cache read costs a tenth of the input rate, and a twentieth on GPT-6.1 Sol and GPT-6.1 Terra. These models also bill the first write into the cache. A 30-minute write costs 1.25 times the input rate, and a 12-hour write set through the Responses API costs 2.5 times. GPT-5.5 and older models charge nothing extra for a write.

A write is not added on top of input. Each input token bills once, at the uncached, cached or cache-write rate. On GPT-5.6 and later, the shortest prefix that can be cached is 1,024 input tokens.

US dollars per million tokens, prompts up to 272,000 input tokens. A 12-hour write costs twice the 30-minute write.

A 30-minute write pays for itself the first time a later request reads the prefix. A 12-hour write needs two reads to come out ahead, so it suits a prompt reused across a long working day.

The Batch API runs requests asynchronously at half the standard input and output rates. It returns results within 24 hours and has its own, higher rate limits. It suits evaluations, classifying large datasets and embedding jobs. The glossary explains prompt caching, cache writes and the batch API in more detail.

OpenAI sells the same models at several speeds, and each speed has its own rate card. Standard is the default and the rate in the price table at the top of this page.

• Flex bills at Batch rates, half of Standard, in exchange for slower responses and occasional unavailability. You choose it per request by setting service_tier to flex. It is in beta on selected models.
• Fast mode costs twice Standard for faster processing. OpenAI renamed it from Priority processing on July 30, 2026.
• Ultrafast costs six times Standard. OpenAI sells it only on GPT-6 Astra and GPT-6.1 Sol, with rate limits of its own.

US dollars per million tokens, input / output, prompts up to 272,000 input tokens. Fast mode is not available for GPT-6 Astra with EU data residency.

Pick a tier by deadline. Flex and Batch cost the same, so Flex suits single requests you can wait on or retry, and Batch suits whole jobs that can wait up to a day. Fast mode and Ultrafast pay off only where response time is worth more than the token bill. The service tiers entry in the glossary compares them across providers.

A prompt longer than 272,000 input tokens moves the whole request to a long-context rate card. Input and cached input double, and output rises by half. GPT-6 Astra goes from $10 to $20 input and from $50 to $75 output per million tokens. GPT-6.1 Sol goes from $2 to $4 input and from $10 to $15 output.

Regional processing endpoints, OpenAI's data residency option, cost 10% more for models released on or after March 5, 2026. FedRAMP endpoints carry the same 10% uplift. The default global endpoints have no uplift.

Built-in tools bill on top of tokens. Web search costs $10 per 1,000 calls, plus the search content tokens at the model's rates. File search costs $2.50 per 1,000 tool calls, and its storage costs $0.10 per GB a day after the first free GB. Hosted Shell and Code Interpreter containers start at $0.03 per 20-minute session for 1 GB of memory.

OpenAI models on Microsoft Azure and Amazon Bedrock are billed by those clouds at their own rates, not at the rates on this page. The long-context pricing entry in the glossary shows how other providers step their rates.

A ChatGPT plan and the OpenAI API are separate products with separate bills. A plan pays for the ChatGPT apps and Codex with a usage allowance. The API bills every token at the rates in the table above and has no monthly fee.

Plan prices from our plan sheet, checked 2026-11-07 on learn.chatgpt.com and help.openai.com. API credit amounts from OpenAI's help center.

The credit is not applied on its own. The plan owner claims it in ChatGPT billing settings or on the API platform and links one API organization they own. Monthly credit expires at the end of the billing cycle and does not roll over. It does not cover Codex run on an API key.

Pay per token when your use is light or uneven. If a month of your work would cost less than $20 at API rates, the API is the cheaper choice. A plan wins when you work in ChatGPT or Codex every day, because its allowance is not metered in dollars. On Pro, the included credit also covers that much API usage each month. The subscription vs API calculator runs the numbers for your own volume.

For the plans side by side, see ChatGPT Go vs Plus and ChatGPT Plus vs Pro.

The OpenAI API is pay as you go. You buy prepaid credit, and your account's usage tier rises with the total you have bought. Each tier sets a monthly usage limit and higher rate limits.

From OpenAI's rate-limit guide.

On Build, GPT-6 Astra, Sol and Terra allow 5,000 requests and 1 million tokens per minute. On Grow, they allow 15,000 requests and 40 million tokens. Batch jobs draw on a separate, larger pool, and Ultrafast has limits of its own.

Some organizations can get free daily tokens by sharing their API traffic with OpenAI for model training. Eligible Launch and Grow organizations get up to 1 million tokens a day on one group of models, which includes GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. They get up to 10 million a day on a second group of smaller models, such as GPT-5.6 Luna and GPT-5.4 mini. Eligible Build organizations get a quarter of each. The account still needs a positive balance, and usage above the quota bills at normal rates.

You create an API key on the OpenAI platform at platform.openai.com. The model ID under each name in the price table is the string you pass as model. When an account runs out of credit, requests fail with HTTP 429. The glossary explains rate limits and free tiers across providers.