OpenAI API pricing: GPT rates per million tokens

As of November 3, 2026, OpenAI API pricing depends on the model, token type and processing tier. For short prompts on Standard, GPT-6.1 Sol costs $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. Batch and Flex each halve those text rates. Batch is asynchronous, while Flex can return capacity errors.

Selected GPT text model prices

OpenAI lists text prices per million tokens. A token is a chunk of text the model reads as input or generates as output, and output often costs more than input.

These tables cover selected GPT models as of November 3, 2026, using OpenAI's pricing page and its model pages. They include realtime text rates and several older and specialized models, but they are not a complete inventory of model IDs, aliases or snapshots. Audio, including realtime audio, image and embedding prices are excluded.

All rates are direct OpenAI Standard prices in USD per million text tokens. Where OpenAI distinguishes context lengths, the tables use short-context rates. "Not listed" means no cached-input price appeared in the cited source; it does not mean caching is free.

Flagship models

For GPT-6 Sol, "Short context" means up to 272,000 input tokens. OpenAI says the GPT-5.6 Sol promotional price is available at least through November 21, 2026.

GPT-6 Sol and gpt-6.1-sol are different models with different cached-input rates. The chat-latest model page says the alias points to the latest Instant model used in ChatGPT, and its underlying snapshot changes regularly.

Older models

A price on this list does not guarantee access. Check the model catalog and your account before choosing an older model. The gpt-4o-2024-05-13 snapshot is omitted because its model page marks it shut down.

Coding models

The model catalog marks gpt-5.3-codex deprecated. Check which coding model your account can use before applying this rate to a new integration.

Pro models

The gpt-5.5-pro model page and gpt-5.4-pro model page provide the Standard input and output rates for those rows. GPT-5.5 Pro's page states that no cached-input discount applies. GPT-5.4 Pro's cached rate is unverified, so the table does not infer one.

For the older rows, see the gpt-5.2-pro model page and OpenAI pricing. The model catalog marks gpt-5-pro deprecated.

Specialized models

OpenAI says gpt-5.6-cyber requires separate approval and provisioning through Daybreak. Access to the Rosalind models is limited to approved internal research through OpenAI's trusted-access program.

The gpt-6-astra-law row is a legal model configuration, and its promotional price is available at least through December 31, 2026. Check OpenAI's specialized pricing and the access conditions before estimating a workload.

Realtime text rates

These prices cover text tokens used with realtime models. Realtime audio has different rates and is excluded from this table.

The 2.1 rates come from OpenAI's realtime pricing. The other rates come from the model pages for gpt-realtime-2, gpt-realtime-1.5, gpt-realtime, gpt-realtime-mini and gpt-4o-realtime-preview.

For prices outside OpenAI, compare the same workload in our LLM API price comparison. A lower token rate is a starting point; token usage and the number of successful answers affect the bill too.

When do cache writes change the price?

Prompt caching lets a model reuse an eligible prefix of input text. For GPT-5.6 and later, OpenAI requires a minimum cacheable prefix of 1,024 visible input tokens, and writing that prefix to the cache can have a separate charge.

The default 30-minute cache write costs 1.25 times the uncached input rate. OpenAI also lists a 12-hour minimum lifetime in the Responses API at 2.5 times that rate. See the prompt caching guide for eligibility and conditions.

These short-context rates come from OpenAI pricing and the GPT-6.1 Sol model page. For GPT-6.1 Sol and Terra, a cache read costs 5% of the ordinary input rate, but the initial write costs more than ordinary input.

Take a 10,000-token eligible GPT-6.1 Sol prefix sent 10 times within the 30-minute cache lifetime. Our assumptions are one cache write followed by nine cached reads, using the conditions in the prompt caching guide.

The write costs (0.01M / 1M) x $2.50 = $0.025, and the nine reads cost (0.09M / 1M) x $0.10 = $0.009. Together, that is $0.034.

Sending the same prefix uncached 10 times costs (0.1M / 1M) x $2 = $0.20. These assumptions save $0.166 on the prefix; other input and output tokens are excluded. For GPT-5.6 and later, explicit-only mode with no explicit breakpoints avoids cache writes. That is how to reproduce the uncached baseline used here.

When do long prompts cost more?

For GPT-6.1 Sol, a prompt over 272,000 input tokens triggers higher rates for the full request. The increase is twice the short-context input and cache rates, and 1.5 times the output rate.

For example, a request with 273,000 input tokens uses the long-context rates for all tokens in the request. Check the model page for the rates that apply to your processing tier.

How much do Batch and Flex save?

For short-context GPT-6.1 Sol, both Batch and Flex charge half the Standard text rates. They are alternative processing tiers, so the two discounts do not stack.

The rates come from the GPT-6.1 Sol model page. OpenAI also lists Fast at twice Standard and Ultrafast at six times Standard for this model.

Batch has a separate rate-limit pool and fits work that can wait for asynchronous results. Its 24-hour completion window does not guarantee that every request succeeds: unfinished requests can expire. See the Batch guide.

Flex uses service_tier: "flex". If capacity is unavailable, OpenAI suggests retrying with Standard when the extra cost is acceptable; see the Flex guide. Check the model's pricing and tier eligibility before applying these rates elsewhere.

What would a workload cost?

Take an illustrative workload of 1,000 successful GPT-6.1 Sol requests. Our assumptions are 1,000 input tokens and 500 output tokens per request, giving 1 million input tokens and 0.5 million output tokens in total.

Each request is below the 1,024-token cacheable-prefix minimum and the long-context threshold, so this example has no cache reads or writes. It is text only and excludes tools, storage, regional uplift, failed retries, taxes and gateway fees. The rates come from the model page; the prefix minimum comes from the prompt caching guide.

On these assumptions, Batch or Flex saves $3.50 on the successful workload. A Flex retry on Standard or an expired Batch request can change both the cost and the time needed to finish the work.

Now, the break-even check for Flex. Assume every initial Flex request incurs the full modeled Flex cost, and a fraction f fails and is retried once on Standard at the full modeled Standard cost.

The cost per completed task is then 0.5 + f times the Standard-only cost. At a 50% retry share, the cost equals Standard; above that share, it is higher. The weakest assumption is the retry rate, and failed-request charges and token counts can change the result.

The cache break-even check uses the GPT-6.1 Sol prefix above. One 30-minute cache write costs 1.25 times an uncached input, or 25% more if the prefix is used once. A second use within the cache lifetime adds a read at 0.05 times the input rate, bringing the total to 1.30 times one uncached use, against 2 times for two uncached uses. The weakest assumption is how often the prefix is reused while it remains eligible for a cache read.

1. 1
   Measure tokens per successful request

   Record input and output tokens for a sample of your own tasks. Keep cached input and cache writes separate where they apply.

2. 2
   Multiply by expected request volume

   Multiply each token count by the number of requests you expect to run, including an estimate for retries.

3. 3
   Apply the model and tier rates

   Divide each token total by one million, multiply by the applicable rate, then add tool and other charges.

What else can appear on the bill?

The text table covers one part of an API estimate. Use the OpenAI pricing page for the model and services you plan to use.

Search

OpenAI lists web search at $10 per 1,000 calls, with search content tokens billed at model rates. Check the pricing page for tool conditions and the model you use: the search call price is separate from the token prices above.

Audio and regional processing

OpenAI lists GPT Transcription at an estimated $0.0045 per minute and GPT-4o Mini Transcription at an estimated $0.003 per minute. Those estimates are separate from the text-token table.

Eligible regional processing endpoints have a 10% uplift for models released on or after March 5, 2026. Check the model and region in the pricing details before adding it.

Is ChatGPT Plus cheaper than the API?

ChatGPT Plus costs $20 per month, and OpenAI API usage is billed separately. Plus does not qualify for the subscription API credit benefit, although other eligible subscriptions may receive credits.

New API accounts use prepaid billing, with a $5 minimum initial purchase. Free credits may be available, but OpenAI does not guarantee them for every new account.

There is no universal price comparison: Plus is a subscription, while an API bill depends on the model and usage. The $7 Standard example above is an API estimate for the stated workload, not a measure of equivalent ChatGPT usage.

How can a team control spend across models?

The lowest token price can lose its appeal if a model needs more attempts to complete the same task. Measure cost per successful task alongside answer quality before moving a workload to a cheaper model.

If you are testing alternatives, Requesty's routing and fallback policies can help you compare models and maintain availability. For the integration path, see one API for GPT, Claude, Gemini and open-weight models.