Updated July 2026. What a single token actually costs across the model lineup, why input and output differ, and how to turn a per-token rate into a real monthly bill.
LLM providers quote prices per million tokens, which hides what a single token costs. This page converts the headline rates into real per-token numbers, explains the input/output split, and shows how to get from cost-per-token to your total bill. To skip the math, drop your prompt into the calculator.
Divide the per-million rate by 1,000,000 and you get the cost of one token:
The numbers are tiny per token, which is exactly why they are quoted per million — and exactly why volume is what turns them into a real bill.
On essentially every model, output costs about 5× input. Claude Opus 4.8 is $5 / $25 per million; Gemini 3.1 Pro is $2 / $12. The reason is compute — every output token is a full forward pass, while input is read in parallel. Full explanation in AI token cost explained.
An app sends 1,500 input tokens and gets 600 output tokens per call, on Gemini 3.1 Pro ($2 / $12 per million):
Swap to Claude Haiku 4.5 ($1 / $5) and the same volume drops to roughly $750/month. Per-token differences look microscopic and add up to thousands. Compare any two models on the compare pages.
You can't negotiate the published rate, but you can lower what you effectively pay:
Per single token it ranges from about $0.0000001 (budget input) to $0.000025 (flagship output). Providers quote per million, and input and output are priced separately.
No — output costs about 5× input on most models. Opus 4.8 is $5 per million input and $25 per million output.
Use prompt caching (~10% on cached input), the Batch API (50% off), a cheaper model, and shorter output.
Multiply token counts by the per-million rate and divide by a million, separately for input and output, then add. The gpt-cost.com calculator does it automatically.
Prices synced from each vendor's official pricing page, July 2026. For exact billing, use each provider's official tokenizer.