TokenAtlasHow OpenAI API cost is calculated
Cost = (input_tokens / 1M) × input_rate + (output_tokens / 1M) × output_rate. GPT-4o sits at $2.50 in / $10 out per 1M tokens, GPT-4.1 at $2.00 in / $8 out, and GPT-3.5-turbo at $0.50 in / $1.50 out. TokenAtlas applies these rates per call and rolls up to monthly volume.
Worked example
A support agent answering 50,000 tickets/month at 800 input + 400 output tokens runs ~40M in / 20M out. On GPT-4o that's $100 + $200 = $300/mo. On GPT-3.5-turbo the same workload costs $20 + $30 = $50/mo — a 6× spread driven entirely by model choice.
What this calculator outputs
Per-request cost, daily burn rate, monthly projection, and the cheapest equivalent model for the same prompt shape. Results are deterministic and exportable.
When to use it
Before shipping a new feature, before signing an enterprise contract, and every time OpenAI revises pricing. Run scenarios on volume, prompt length, and output length to find the safe operating envelope.
Cost drivers ranked
For a given OpenAI workload, output length matters most because output tokens are billed 3-4x input on GPT-4o and GPT-4.1. Next is call volume, since cost scales linearly with monthly calls. Prompt length ranks third, and retried or failed calls rank fourth but are easy to miss: a 5% retry rate adds 5% to the modelled bill even though no new user request was served.
Choosing a model for the workload
Blended at 70% input / 30% output: GPT-4o = 0.7x$2.50 + 0.3x$10 = $4.75/1M tokens. GPT-4.1 = 0.7x$2 + 0.3x$8 = $3.80/1M. GPT-4o mini = 0.7x$0.15 + 0.3x$0.60 = $0.285/1M. If the task tolerates a smaller model, mini is roughly 17x cheaper than GPT-4o on this blend — the calculator lets you test that swap against the same volume.
What moves the estimate most
Three inputs swing an OpenAI estimate more than any tuning of the model picker: whether the system prompt is re-sent every turn (multi-turn chat re-bills it each call), whether tool calls add 100-300 tokens per invocation, and whether output length is capped. Re-run the same volume with each of these toggled to see which one actually drives the monthly number.
Reading the result correctly
The output is a modelled estimate built from the token volumes and rates you entered — it is not a bill and it does not read your OpenAI account. Write down the assumptions that produced it (prompt size, completion size, calls/month) alongside the number, since the estimate is only as good as those inputs and will need re-running if any of them change.
Frequently asked questions
- Are the rates current?
- Yes — TokenAtlas tracks OpenAI's published pricing and updates within 24h of a rate change.
- Does it cover fine-tuned models?
- Yes. Fine-tuned GPT-4o and GPT-3.5 are priced separately and included in the model picker.
- Can I export the result?
- PDF and CSV export are available on the Pro plan.
- Does it include cached input pricing?
- Yes. GPT-4o cached input is billed at 50% and applied automatically when you toggle 'prompt caching'.