How much does the OpenAI API cost for a small team?

The OpenAI API has no single monthly price. A team pays for what its software sends and receives, using rates that vary by model. Build an estimate from input, cached input and output tokens, then add any tool or storage charges. ChatGPT subscriptions are billed separately and do not include API use.

The OpenAI API separates token prices

An API, or application programming interface, lets software send requests to a model without a person using the ChatGPT website. OpenAI’s current pricing table gives each text model three standard rates, all in dollars per one million tokens. A token is a small piece of text processed by the model, explained in our guide to AI tokens.

Input tokens are the instructions, conversation and documents sent to the model. Cached input tokens are repeated input that OpenAI can reuse from a recent request at a lower rate. Output tokens are the answer the model generates. On the standard table fetched on 7 September 2026, GPT-6 Astra is $10 per million input tokens, $1 per million cached input tokens and $50 per million output tokens. GPT-5.6 Luna is $0.20, $0.02 and $1.20 on those same three lines. The chart shows every OpenAI model currently stored in our pricing ledger, not every historical model on OpenAI’s page.

Those model rows make model choice visible as a budget decision. The Astra input rate is fifty times the Luna rate, while both are sold by OpenAI. A higher-priced model may solve a difficult task in fewer attempts, but its name alone does not prove that. Test the cheapest model that meets a written quality requirement, then move up only where the results justify it.

A small-team estimate, step by step

Start with one repeatable workflow, not the whole company. Suppose a support tool handles 10,000 requests a month. Each request sends 2,000 input tokens and receives 400 output tokens. That is 20 million input tokens and 4 million output tokens. If none of the input is cached, the monthly model cost is:

20 x input rate + 4 x output rate

At the tracked GPT-5.6 Luna standard rates, that arithmetic is 20 times $0.20 plus 4 times $1.20, which equals $8.80. At GPT-6 Astra’s standard rates, it is 20 times $10 plus 4 times $50, which equals $400. These are worked estimates from the published rate card, not quotes. They exclude retries, hidden instructions, tool calls, long-context pricing, storage and taxes.

Now measure a real pilot. OpenAI returns usage metadata, the token counts attached to a response. Record the input, cached input and output fields by workflow and model. Use a median for typical use and a high percentile for busy or unusually long requests. A percentile is a point in the distribution: the 95th percentile is larger than 95 percent of observed requests. This prevents one small demo from becoming the budget for a messier production month.

Caching can materially change repeated work. OpenAI’s prompt caching guide says eligible repeated prefixes are cached automatically. A prefix is the unchanged beginning of a request, such as a long policy document followed by a new question. The response usage data reports cached tokens, so a team can verify the saving rather than assume it. Keep stable instructions at the start and changing material later when that also produces correct results.

The charges outside the token table

The main table is not the whole price page. OpenAI publishes separate charges for built-in tools and storage. A built-in tool is a vendor-operated capability, such as web search or file search, that a model can call during a response. The unit may be a tool call, stored gigabyte-day or additional content tokens. Check the exact tool row used by the application before multiplying model tokens.

Service choice matters too. Batch processing groups requests for asynchronous completion rather than immediate replies. OpenAI’s cost guide points to Batch for workloads that can wait, and the pricing page has separate Batch rates. Latency is the wait between a request and its result. Interactive support needs low latency; an overnight classification job may not. Do not pay for speed the workflow cannot use.

Long inputs can have special pricing. The current OpenAI table separates standard and long-context rates for supported models. Long context means a request whose input passes the model’s stated threshold. For GPT-6 Astra, the standard pricing page shows higher long-context columns. Our guide to context windows explains why the maximum advertised window and the normal input rate should never be multiplied without checking that condition.

Also budget for failed and repeated work. A timeout in your app may trigger a retry, and a poor prompt may produce a second full answer. Track requests with a stable identifier so retries are visible. Testing prompt quality can reduce cost more safely than deleting context the model needs.

Budgets, alerts and the limit that actually stops work

A rate limit controls speed, such as requests or tokens per minute. It protects capacity and can return an error when traffic arrives too quickly. A spend limit controls accumulated cost. Each control answers a different question, so raising a rate limit does not create a budget and lowering a budget does not guarantee smooth traffic.

OpenAI’s Help Center distinguishes spend controls from notification thresholds. A notification threshold sends an alert when usage crosses a chosen amount. It does not necessarily stop later requests. Where the account exposes hard project or organization limits, use those for a ceiling, and keep alerts below the ceiling so someone has time to investigate. Give experiments their own project rather than mixing them with production. That produces separate usage views and limits the effect of a runaway test.

OpenAI also states that ChatGPT and the API platform have separate billing systems. The consumer ChatGPT plans page can help a person choose a chat subscription, but it is not the rate card for an application. Compare APIs on the vendor pricing page and consumer products on the plan page.

What it means for you

For a small team, start with columns for workflow, model, monthly input tokens and monthly output tokens. Add cached input only after a pilot shows it in usage metadata. Put tools and storage in separate rows. Then run the estimate at both typical and busy-month volumes, with a contingency for retries.

Choose a model with a test set that reflects the work. A test set is a fixed collection of real examples scored by the team against a clear standard. Price every model that passes, not every model that exists. Review the ledger date and OpenAI’s source page before approval because rates and model names can change.

Finally, make ownership visible. One person should receive spend alerts, one hard limit should cap each project, and the invoice should map back to logged workflows. That turns OpenAI API pricing from an unknowable monthly surprise into multiplication your team can audit.