Input and output token prices, explained
Input and output tokens carry separate per-million rates, output is commonly dearer, and a blended rate only means something for one fixed mix of the two.
Two rates, and why the second is usually higher
An LLM API charges one rate per million tokens you send in and another per million tokens the model writes back, and the output rate is commonly the higher one. A common explanation is mechanical: input can be processed in one parallel pass, while output is produced a token at a time, each step waiting on the last, which keeps the hardware busy longer per token. Your bill depends on both rates and on the ratio between your input and output.
The token price table shows both rates for each listing (input-price listings on record: 750; providers: 16). The calculator does not cost a listing that carries only one half of the price, because half a price produces a confident wrong total.
Free tiers and promotional rates complicate the picture. A seller may list a free row for a model with tight limits beside a paid row for the same model. Those are different products, and a comparison that mixes them answers neither question, so compare paid rows with paid rows.
Your traffic has a shape
Many workloads lean hard one way. Summarising a long contract is mostly input: thousands of tokens in, a paragraph out. Drafting marketing copy from a one-line brief is the reverse. Classification sends a document and gets back a single label, so output barely registers.
Agents and chat sit in the middle and drift over time. Each step of an agent resends the growing history as input and writes a short tool call as output, so the input side swells with the length of the task while the output side stays modest.
Before comparing prices, write down your own ratio. Average output tokens per request divided by average input tokens per request is the number that decides which rate matters more to you.
What a blended rate is and where it misleads
A blended rate folds the two prices into one figure for a stated mix. The method: multiply the input rate by the input share of your tokens, multiply the output rate by the output share, and add them. The result is a cost per million tokens of all kinds, for that mix and no other.
Published blended figures assume somebody else's mix. A comparison blended at mostly input, read by someone whose workload is mostly output, can show a ranking that flips once the reader's own shares go in. Models that look close on a blended figure can be far apart on the half of the price your traffic actually uses.
So blend only with your own ratio, and keep the two raw rates beside the blended one. The raw rates are what the seller published. The blended figure is your arithmetic, and it is only as good as the ratio you fed it.
Pricing a contract summariser
Take a tool that reads uploaded contracts and returns a short summary. Monthly input is documents per month times average tokens per document, plus the instructions on each call. Monthly output is documents per month times average summary length. Each total is divided by a million and multiplied by its own rate.
Because the input side is so much larger, the input rate dominates this bill, and a seller with a low output rate but a high input rate would cost more here than its headline suggests. Flip the workload to drafting long replies from short prompts and the same seller could come out ahead.
The LLM cost calculator takes the two monthly totals on separate sliders, so the mix is yours rather than an assumption baked into a single figure.
Moving the expensive half
Output is the side you control most directly. A cap on maximum output tokens, an instruction to answer in a sentence, or a structured format that leaves out prose all shorten what the model writes. On a reasoning model, check whether thinking tokens count as output, because they can outnumber the visible answer.
Input is harder to cut once a product depends on context, but it often carries waste: retrieved passages nobody reads, a full history where a summary would do, examples that no longer change the result. Trim those and measure the token counts the API reports before and after.
Model choice moves both halves at once. A smaller model on the same task often carries lower rates on both sides, and the question of whether it answers well enough is yours to settle by testing. Nothing on this site ranks models on quality, so the price comparison only starts once you have a shortlist you trust.
When the cheaper rate is the dearer bill
A seller with the lowest input rate on a model can still produce the highest bill for a workload that is mostly output. The reverse also happens. The only way to know is to price your own month at each seller, with your own ratio, on both rates at once.
The LLM price index shows how token rates sit relative to each other. It cannot tell you which seller is cheapest for you, because it does not know your mix; the calculator does that step once you give it your two monthly totals.
Keep a note of the ratio you used and the date. When a seller changes one of its rates, rerun the sum with the same ratio rather than reading the new headline figure, so the comparison stays like for like.
What we could not answer
- ⚑We hold list rates, not the ratio of input to output in anyone's real traffic, so the shapes described here are typical patterns rather than measurements.
- ⚑The reason output is dearer is general serving mechanics; a seller's own cost structure is not something we record.
What this rests on
infratoapp, “Input and output token prices, explained”, read 27 October 2026. input-price listings on record: 750. https://www.infratoapp.com/blog/input-and-output-token-prices-explained
Questions people ask about this
Why do output tokens cost more than input tokens?
Input is processed in a single parallel pass, while output is generated one token at a time, each depending on the one before. That is a common explanation for output being priced higher, though a seller's own reasons are not on record. Not every seller prices it higher, so read both rates for the model you use.
What is a blended token price?
It is a single cost per million tokens that combines the input and output rates for a stated mix, such as mostly input with some output. It is useful only for that mix. If your own ratio differs, recompute it with your shares or compare the two raw rates directly.
Which rate matters more for my workload?
Divide your average output tokens per request by your average input tokens per request. A small result, as with summarising or classifying, means the input rate dominates. A large one, as with drafting long text from short prompts, means the output rate dominates the bill.
These render as FAQPage structured data · answers are filled from the same live figures
Read next
A token is a chunk of text, about four characters of English, and an LLM API bills every token it reads and every token it writes at separate rates.
A cached input rate bills a repeated prompt prefix for less, but only where the seller publishes one; everywhere else the same tokens pay the full input rate.
A large context window is a limit, not a charge; the charge is every token you actually place in it, billed as input on every call that carries it.