The formula
Model.sale charges one flat rate per million charged tokens. Charged tokens = fresh input + output + 0.4 x cached input, because cached input is billed at 40%. The cost of a request is charged tokens divided by 1,000,000 times the model's price. There is no per-request fee, no subscription and no minimum beyond the first top-up.
The live prices at the time of writing: gpt-6-luna $0.08, gpt-6-sol $0.38, gpt-6.1-sol $0.40 and gpt-6-astra $1.125 per million charged tokens. The current numbers are always on the models page; prices can change, and a request is charged at the price in force when it ran.
A chat message
A typical chat turn is a short question and a medium answer: about 500 tokens in and 700 out, so 1,200 charged tokens. On gpt-6-sol that is 1,200 / 1,000,000 x $0.38 = $0.00046, so a thousand such messages cost about $0.46. On gpt-6-luna the same thousand cost about $0.10. A long conversation that re-sends 10,000 tokens of history on every turn is a different case: the history is mostly cached input, counted at 40%, which keeps the total far lower than the raw token count suggests.
An API call in your product
Take a support-reply feature that sends 1,500 tokens of instructions and context and gets 300 back: 1,800 charged tokens. On gpt-6-sol, 100 requests cost 100 x 1,800 / 1,000,000 x $0.38 = $0.068, and 100,000 requests a month cost about $68. Moving the easy requests to gpt-6-luna would cut that by roughly four fifths. Summaries, classification and extraction rarely need the biggest model.
Write your own version of this calculation with real numbers from a test run: the dashboard shows the exact charge and token counts of every request, so measure ten typical requests and multiply.
A coding session
Agents behave differently from chat. A medium refactoring session may send 150,000 fresh input tokens, read 2,000,000 tokens from cache across many steps and produce 60,000 output tokens. Charged tokens = 150,000 + 60,000 + 0.4 x 2,000,000 = 1,010,000. On gpt-6-sol that is about $0.38; on gpt-6-astra about $1.14. The Codex CLI cost guide shows the same example step by step, and the Cursor and Codex pages explain how to connect them.
The cache is why long sessions stay affordable, and also why a very rough guess based on the raw token count overestimates the bill by a large factor.
What moves the bill
The model is the biggest lever: Luna to Astra is about fourteen times in price per charged token. Start with Luna to test an integration, run daily work on Sol and move a task to Astra only after it fails twice. The second lever is output length: set a maximum output so a runaway answer cannot grow the request. The third is repeated context: keep the stable part of your prompt at the start so it is served from cache.
Retries also cost: every attempt that produces output is charged for the usage it reports, so cap automatic retries in your client and keep an eye on an agent that loops.
Budget before you pay
The balance is prepaid, so spend cannot exceed what you add. Create a key per project, give it a daily limit and an allowed-model list, and top up from $5. Use the cost calculator on the site to estimate a workload before you run it, and check the first real request in the dashboard against your estimate. If they match, scale up.