OpenAI Cost Calculator

Three hundred dollars a month at a thousand requests a day

At 1,000 requests a day with 1,000 input tokens and 500 output tokens each, the projected cost is $10 a day, $300 a month, and $3,650 a year — $0.01 per request.

A cent per request sounds negligible. At scale it is not, and the arithmetic is worth doing before a feature ships rather than after the first invoice.

How the projection is built

Cost per request = (input tokens × input rate) + (output tokens × output rate)

Monthly = cost per request × requests per day × 30

The three inputs that drive it — request volume, input length and output length — should all be measured rather than guessed. The usual pattern is that input tokens are underestimated, because system prompts, retrieved context and conversation history are forgotten, and output tokens are underestimated because models are verbose unless constrained.

Where the cost actually goes in real applications

Conversation history is the most common surprise. In a chat interface, every turn resends the entire prior exchange as input. A twenty-turn conversation sends roughly twenty times the context of the first turn, so total cost grows with the square of conversation length rather than linearly. Truncating or summarising history is not a refinement — it is the difference between a sustainable and an unsustainable bill.

Retrieval-augmented generation attaches retrieved documents to every prompt. Ten retrieved chunks of 500 tokens each add 5,000 input tokens per request, which typically dwarfs the user's actual question.

System prompts are sent on every single request. A 2,000-token system prompt at a million requests a month is two billion input tokens of pure fixed overhead — and it is the ideal candidate for caching.

Retries and failures. Requests that error or return unusable output are still billed.

The levers, roughly in order of impact

Prompt caching. Where a substantial prefix is identical across requests, cached input is billed at a heavy discount. For applications with long system prompts or fixed document context, savings of 50–90% on the input portion are achievable.

Batch processing. Where results are not needed immediately, batch APIs typically offer around 50% off for asynchronous completion within a window. Well suited to bulk classification, enrichment and offline generation.

Model routing. Send simple requests to a small model and escalate only when needed. A classifier deciding which model to use costs almost nothing and can cut total spend dramatically.

Output constraints. Specify maximum length, request structured output, ask for bullets rather than prose. Output is the expensive half.

Caching complete responses. Many applications receive the same question repeatedly. A conventional response cache keyed on the normalised prompt eliminates the call entirely.

Reading the result

The figure is a steady-state estimate and does not model the spikes that usually cause budget overruns — a viral moment, a misbehaving client retrying in a loop, or an automated job with no rate limit.

Set hard spend limits at the provider level and rate limits per user before launch rather than after. An unbounded loop against a paid API is a well-documented way to generate a very large bill overnight.

Note also that the annual figure uses 365 days while the monthly uses 30, so the annual is not simply twelve times the monthly.

Common mistakes

Forgetting the system prompt. It is charged on every request.

Ignoring conversation growth. Chat cost scales quadratically without truncation.

Projecting from a single test call. Measure across realistic usage.

Omitting retries. Failed calls are billed.

Launching without a spend cap. The most expensive mistake on this list.