How to read an AI price list
An AI price list starts with two numbers per model: a price to read and a price to write. This guide takes one job, prices it on four models, then shows how reasoning, caching and batching move the bill.
The headline prices for an AI model are two numbers, both quoted per million tokens. GPT-5.5 is listed at $5 for input and $30 for output. Other rates sit beside them on the same page: a cheaper rate for cached input, a discount for batches, and sometimes a higher rate for very long prompts. This guide explains the parts of a price list in the order they affect what you pay.
1. Tokens, not words
Models count text in tokens, which are chunks of a word. OpenAI's guide to tokens (read on 14 November 2026) gives the rule of thumb that 100 tokens is about 75 words of English. So a million tokens is roughly 750,000 words, and a 1,500-word report is about 2,000 tokens.
The rule is for English prose. Code, tables and other languages come out differently, so count a real sample of your own text with a token counter before you trust an estimate.
2. Input and output are priced apart
The input price covers everything you send: your instructions, any documents, and the whole conversation so far. The output price covers what the model writes back. Output is usually dearer. Among models released since July, the typical output price in this snapshot is 5 times the input price, compared with three times for late-2024 releases. These are release cohorts priced on the snapshot date, not historical price observations. We looked at that in a separate article.
A chat application that keeps the full history does not send only your latest message. It sends the whole conversation again on every turn, so the input side of a long chat grows with each reply. Applications can trim or summarise the history to limit that.
3. One job on four models
Here is an illustrative job: 10,000 requests in a month, each sending 3,000 tokens and getting 500 back. That is 30 million input tokens and 5 million output tokens. The token counts are assumptions to show the sums. The prices are real.
| Model | Price from | Input price | Output price | Input cost | Output cost | Month's bill |
|---|---|---|---|---|---|---|
| Mistral Nemo (Mistral) | OpenRouter | $0.029 | $0.03 | $0.87 | $0.15 | $1.02 |
| Claude Haiku 4.5 (Anthropic) | OpenRouter | $1 | $5 | $30 | $25 | $55 |
| GPT-5.5 (OpenAI) | Maker's list | $5 | $30 | $150 | $150 | $300 |
| GPT-5.5 Pro (OpenAI) | Maker's list | $30 | $180 | $900 | $900 | $1,800 |
The same month of work on four models
The same work costs about a dollar or about $1,800, a gap of roughly 1,765 to one. That is smaller than the 6,000-fold gap in output prices that we wrote about in another piece, because this job is mostly input and input prices sit closer together. Notice too that on GPT-5.5 the output side costs as much as the input side, even though the job sends six times more tokens than it receives.
4. The other rates on the list
Now keep the model fixed at GPT-5.5 and change how the job is run. The list price gives $300 a month.
| How the job is run | Month's bill | Against list |
|---|---|---|
| As listed Input and output at list price | $300 | n/a |
| With a cached opening Best case: 2,560 of the 3,000 input tokens repeat, and every request reads them from the cache at $0.50 per million | $184.80 | −38.4% |
| Sent as a batch 50% discount, 24-hour completion window | $150 | −50% |
| With reasoning switched on 2,000 reasoning tokens per request, billed as output | $900 | +200% |
Reasoning. Some models can think before they answer. OpenAI's documentation (read on 14 November 2026) says those thinking tokens are not shown to you but "are billed as output tokens", and that one response can use tens of thousands. In the table, 2,000 of them per request triple the bill. If you switch a reasoning mode on, measure it on a few real requests before you budget.
Caching. If every request starts with the same long instructions or the same document, the provider can keep that opening ready and charge much less to read it back. OpenAI's caching guide says this is on by default for supported models, and its GPT-5.5 page lists cached input at $0.50 per million tokens, a tenth of the normal $5. Anthropic's caching guide gives a standard cached-read price of a tenth of base input, with lower rates on selected models. Writing to its cache costs 25% more than normal input at the default five-minute lifetime, and that lifetime restarts each time the cached text is used. All three pages were read on 14 November 2026. Caching helps only when the repeated text comes first and the next request arrives before the cache expires.
Batch. If you do not need the answer at once, you can hand over a file of requests and collect the results later. OpenAI's Batch API guide (read on 14 November 2026) gives a 50% discount for this, with a 24-hour completion window. Requests not finished in that window are cancelled, and you pay only for the ones that completed. It suits overnight jobs such as tagging a back catalogue. It is no use for a chat window.
5. Whose price is it?
The last thing to check is who is quoting. A model that anyone may host, such as Mistral Nemo or the Llama family, is sold by many hosting companies at different prices, and our tables show the cheapest one that OpenRouter lists. A model sold only by its maker is priced on the maker's own page, and that page can hold more than one rate: OpenAI's GPT-5.5 page lists a higher price for prompts over 272,000 input tokens and a 10% uplift for regional processing. Our own tables mix the two kinds of price, and the first table in this guide shows which is which. Our transparency page explains where each price comes from.
A single "blended" price per million tokens is a third thing again. It assumes a fixed mix of input and output. Ours assumes one part input to three parts output, and the formula is published. The example job in this guide is six parts input to one part output. Apply our blended price to it on GPT-5.5 and you get $831.25, against the $300 worked out above.
A short checklist
- Count the tokens in a real request and a real answer. Do not guess from word counts alone.
- Multiply by your monthly volume, input and output separately.
- If you use a reasoning mode, add its tokens to the output side.
- If your requests share a long opening, check the provider's caching rules and price the repeated part at the cached rate.
- If the work can wait a day, price it as a batch.
- Check whether the price is the maker's list price or a reseller's.
- Run a small real test and read the usage report before you commit.
The cost calculator estimates a monthly cost from the token counts you enter, for the models it supports, and the pricing guide goes over the basics.