ModelRefs / Cost Per Token — AI Glossary
Cost Per Token — AI Glossary
The unit price charged by an LLM API per input or output token; the primary metric for comparing inference pricing across providers. 25/MTok, output $5/MTok.
Overview
Token pricing is asymmetric: output tokens cost 3–5× input tokens (more compute intensive). As of 2025: GPT-4o input $2.50/MTok, output $10/MTok; Claude Sonnet 4 input $3/MTok, output $15/MTok; Gemini 1.5 Pro input $1.25/MTok, output $5/MTok. Effective cost depends on average prompt/completion lengths in your workload.
Reference details
| Topic | ecosystem |
|---|---|
| Also known as | token price, price per token |
| Last reviewed | 2026-06-24 |
Related terms
How it works
- Input and output tokens are priced separately, and output costs several times input because generating is sequential while prefill is parallel.
- A request's cost is therefore the input length times the input rate plus the output length times the output rate — two terms that behave very differently.
- Prompt length is under the caller's control and multiplies by the cheaper rate; output length multiplies by the dearer one.
- Cached input, where supported, is billed at a reduced rate, which changes the arithmetic for repeated long prefixes.
- Reasoning tokens are billed as output whether or not they are shown, so the visible answer understates the cost of a reasoning model.
Example: Which term dominates, and why intuition gets it wrong
Take input at £2 and output at £10 per million tokens — a 5x asymmetry. A request with a 4,000-token prompt and a 200-token answer costs 4,000 x 2 + 200 x 10 = 8,000 + 2,000 = 10,000 units, so input is 80% of it. A request with a 500-token prompt and a 2,000-token answer costs 1,000 + 20,000 = 21,000 units, where output is 95%. Same price list, opposite optimisation: trim the prompt in the first case, constrain the answer in the second. Averaging the two into one blended rate hides which lever applies.
Commonly confused with
Frequently treated as the comparison between providers, when it is one input to it. A lower per-token price can cost more in practice if the model needs more output tokens to reach the same answer, uses a tokeniser that splits your text less efficiently, or requires retries. The comparable unit is cost per completed task, not cost per token. Tokeniser differences alone can move the effective price by a factor on non-English text, which is why a headline rate is not portable between providers.
When to use it
Reach for it when:
- Modelling spend for a workload whose input and output lengths are known.
- Deciding whether to trim prompts or constrain output, which the two terms answer differently.
- Evaluating whether prompt caching repays its complexity for a long reused prefix.
Reach for something else when:
- Comparing providers directly, where cost per completed task is the comparable unit.
- Reasoning models, where hidden thinking tokens make the visible answer a poor cost proxy.
- As a blended single rate, which conceals whether input or output dominates.
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Cost Per Token — AI Glossary.
Frequently asked questions
What is Cost Per Token?
The unit price charged by an LLM API per input or output token; the primary metric for comparing inference pricing across providers.
Is Cost Per Token the same as token price?
Yes — token price, price per token are common aliases for Cost Per Token.
What concepts are related to Cost Per Token?
Closely related concepts include cost per request, tokens per minute, inference cost.