LLM API cost glossary
Short answer: An LLM API bill multiplies a token count by a rate per 1M tokens, and caching, batch, reasoning and context length change either the count or the rate. These are the terms this site uses, each linked to the page where the number is calculated.
The words on an API bill are not hard, but they get used loosely. Each definition below fixes what a term means on this site and links the page where the number behind it is worked out.
Tokens and counting
Token
The unit of text an API counts and bills. English prose averages about four characters per token, so 1,000 tokens is roughly 750 words; code, JSON and non-Latin scripts use more tokens per word. Every rate on this site is quoted per 1M tokens.
Tokenizer
The program that splits text into tokens before the model sees it. Each provider and model generation has its own, so the same paragraph can be a different count on GPT, Claude and Gemini, and a count measured on one model does not transfer to another.
Input tokens
Everything sent in a request: system prompt, tool definitions, retrieved documents, earlier turns and the new message. Input is the cheaper side of most rate cards, but with a large context it is usually the biggest line on the bill.
Output tokens
Everything the model writes, including the hidden thinking on reasoning models. Output rates run from 1.0x to 8.3x the input rate (median 5.0x across 55 models), so answer length moves the bill more than an equal change in prompt length.
Context window
The maximum number of tokens one request can hold, input and output together. Go past it and the request is rejected or the oldest text is dropped, which is why long-document and RAG work needs a window large enough for the retrieved context plus the answer.
Reasoning (thinking) tokens
Tokens a reasoning model generates internally before it writes the visible answer. Providers bill them at the output rate, so a short answer can cost several times its visible length; the reasoning-token guide prices that multiplier per model.
Completion
The older name for output tokens, and the field name in OpenAI-compatible APIs (completion_tokens). Usage reports usually return prompt_tokens and completion_tokens; add both when you check a request against a rate card.
Rates and discounts
Per 1M tokens
How list prices are quoted: dollars per one million tokens. A model at $3 per 1M input tokens charges $0.000003 per input token, so a 2,000-token prompt costs $0.006 before any output.
Standard (list) rate
The published rate for a normal synchronous request, before caching, batch or volume agreements. Every table here starts from the standard rate and turns on another tier only where the provider publishes one.
Cached input
A repeated prefix the provider keeps in cache and bills below the standard input rate. It pays off when the same system prompt or document is sent again inside the cache window; the caching guide works out the break-even.
Cache write
The charge for putting a prefix into the cache. Some providers bill the write above the standard input rate, so caching a prefix that is read back only once can cost more than not caching it.
Batch discount
The lower rate for asynchronous requests the provider can schedule in bulk: about 50% off on OpenAI, Anthropic and Google, and 20% on the xAI models that list one. It suits work nobody is waiting on; the batch guide covers when it applies.
Offline rate
A second, lower rate that some video models list for offline or non-realtime inference, usually half the online rate. Seedance 1.0 Pro and Pro Fast publish one, and the Seedance pricing page prices clips at both.
Long-context tier
A higher rate above a published context threshold. The model page lists the tiers when a provider charges one, and the calculator applies the standard rate unless you enter a cached share.
Image tokens
Images are converted into tokens before a vision model prices them, sometimes per tile and sometimes at a fixed count per image. The Nano Banana Pro cost page turns image-token rates into dollars per image.
Video tokens
BytePlus bills video in video tokens estimated as (input seconds + output seconds) x width x height x frame rate / 1024. The Seedance calculator applies that formula at each resolution.
Credits
Subscription products such as Google Flow sell credits instead of tokens, and a credit buys a different number of clips on each plan. The Flow credit calculator converts credits into dollars per clip first, then compares plans.
Turning rates into a bill
Cost per call
Input tokens times the input rate plus output tokens times the output rate, divided by one million. It is the number to quote when two models do the same job, and the calculator works it out for every verified model.
Blended price
One rate that averages input and output at a chosen mix, usually 3:1 or 4:1. Useful for ranking models in a single column, misleading for one specific workload; the blended-cost page lets you set the mix yourself.
Cost per 1,000 calls
Cost per call times 1,000. A $0.002 call is $2 per 1,000 calls and $2,000 per million, which is the right size of number for deciding whether a feature can ship.
Cost per user per month
Calls per user per day, times 30, times cost per call. It is the input to pricing: divide it by (1 minus the target margin) and you have the lowest subscription price that keeps the margin. The unit-economics calculator runs it with your own counts.
Gross margin
Money left from revenue after the model bill, divided by revenue. A 20% rise in the model rate cuts straight into that margin, which is why the price floor is recalculated whenever a provider changes a rate.
Unit economics
Cost and price per one unit of the business, such as one user, one task or one API call, instead of totals. They say whether growth adds profit or loss before you spend anything on acquisition.
Break-even
The usage level where two options cost the same: a self-hosted GPU against the API, or an API bill against a subscription. The self-host calculator and the API vs subscription page solve for it.
Cost per task
The cost of finishing one job rather than making one call. Agent tasks carry several tool rounds, retries and a growing context, so the task costs a multiple of a single call; the agent cost calculator prices it per completed task.
Minimum price per user
Cost per user divided by (1 minus the target gross margin). It is the floor a subscription needs to hold the margin, not a recommendation for the price to charge; the market decides the part above the floor.
Verification terms
Check date
The day we opened the provider's own pricing page and read the row. Each model page carries its own, and a rate with an old check date should be treated as stale rather than current.
Unverified
A model we know exists but whose official page published no readable rate when we looked. It stays out of the rankings instead of getting an estimated price; the method page states the rule.
Price change
A dated change to a provider's list price: an increase, a promotional window ending, or an intro rate that became standard. The price-change log records the ones providers announce in advance.
Effective date
The date a new rate starts to apply, which can be later than the announcement. Budgets should switch on that date; a price shown today but effective next month is a common way for a forecast to miss.
Where these terms are used
- Every model price per 1M tokens, with check dates
- Cheapest LLM APIs ranked by cost per call
- Announced price changes
- New models and APIs
- How prices are verified
Questions
What is the difference between a token and a word?
A word is a spelling unit and a token is a billing unit. English averages about 0.75 words per token, so 1,000 words is roughly 1,300 tokens, but code and non-Latin text differ enough that only the provider tokenizer gives the real count.
Why is output billed higher than input?
Generating tokens is sequential work, so it costs the provider more than reading a prompt, and the rate card reflects that. On the models tracked here the output rate runs from 1.0x to 8.3x the input rate (median 5.0x across 55 models).
Does prompt caching always save money?
No. It saves when the same prefix is reused inside the cache window and can cost more when it is read back once, because some providers charge a cache write premium. Compare the write cost with the read discount before switching it on.
Is cost per call the same as cost per task?
No. A call is one request; a task can take several calls, tool rounds and retries. Price agent work per completed task, and use cost per call only for single-request features such as classification or summarisation.
How current are the numbers on this page?
This page defines terms, so it carries no prices. The ratio quoted in the output-token definition is computed from the price file at build time, whose newest check date is 2026-11-09; every price to act on sits on a model page with its own check date and source link.
Next step → Turn these rates into a price per user
Price data last checked 2026-11-18 against official sources · how we verify
Price-change alerts
One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.