OpenAI Cost Calculator

The fastest way to understand OpenAI pricing is to run a concrete scenario. Take a customer-support chatbot handling 10,000 conversations a month, with about eight exchanges per conversation. Each request carries roughly 1,500 input tokens — the system prompt plus the running conversation history — and produces around 200 output tokens. That adds up to 80,000 requests, or about 120 million input and 16 million output tokens per month.

On GPT-5.6 Luna, at the rates shown in the calculator above, that workload lands in the low hundreds of dollars per month. Run the exact same traffic through the flagship Sol tier and the bill roughly quintuples, well past a thousand dollars. Nothing about the product changed — only the model did. Model choice, not traffic volume, is usually the single biggest lever on an OpenAI invoice.

Document pipelines behave differently: summarization jobs are input-heavy — long documents in, short summaries out — so they favor models with cheap input rates and benefit disproportionately from batch discounts. For the full scenario math on chat workloads, see how much a chatbot really costs to run; if you're weighing the API against a ChatGPT Plus seat, the breakeven is worked out in API vs subscription.

Chat applications resend the entire conversation history with every turn, so input usage grows with the square of conversation length: turn ten costs far more than turn two, even though the user typed a similar message. Long system prompts compound the same way — a 700-token system prompt is billed on every single request, which across a million requests means 700 million input tokens spent before any user text is even processed.

Verbose outputs are the other silent cost: models default to polite preambles and richly formatted JSON unless you instruct otherwise, and every one of those tokens is billed at the higher output rate. On the flip side, OpenAI discounts cached input for repeated prompt prefixes and charges roughly half price through the Batch API for jobs that can wait — leaving those unused is a real loss.

Most teams over-buy intelligence. A practical pattern is model routing: send easy, high-volume requests to a small model and escalate only ambiguous cases to a larger one. Combine that with three habits — trim the system prompt to what the model actually needs, cap or summarize conversation history after a few turns, and set explicit limits on response length.

Before committing, measure instead of guessing: paste a representative prompt into the token calculator to see what it really costs across models, and use the GPT vs Claude comparison to check whether an equivalent Anthropic model prices better for your input/output mix.

List prices only mean something next to the alternatives, so here is the August 2026 table the calculator uses, per million tokens (input / output). OpenAI runs three tiers on the same 1.05M-token context: GPT-5.6 Sol at $5.00 / $30.00, Terra at $2.00 / $12.00 and Luna at $0.20 / $1.20. Terra and Luna were repriced on 30 July 2026 — Luna by 80% and Terra by 20% — so any quote you took in July is now too high. Anthropic's line-up sits close to Sol on input and cheaper on output — Claude Opus 4.8 $5.00 / $25.00, Sonnet 5 $3.00 / $15.00 and Haiku 4.5 $1.00 / $5.00 — but with a 200k window instead of a million.

That repricing moved Luna into territory that used to belong to the open-weight models. Gemini 2.5 Flash costs $0.30 / $2.50, Nemotron 3 Ultra $0.50 / $2.20, and Llama 4 Maverick $0.20 / $0.60 — Maverick is still the floor, but Luna now undercuts Flash and Nemotron on both sides of the meter while carrying a frontier-family name. Against Sol, Luna is 25× cheaper on output for the same context window. Whether the capability trade is worth it depends entirely on the task; the honest way to find out is to price the same workload on both and then run your own evaluation. Our GPT vs Claude comparison does the head-to-head, and the cheapest LLM API guide covers the budget end.