How AI token pricing works
Why AI models charge per million tokens, why output costs more than input, a worked cost example with real prices, and how caching and batch discounts change the bill.
Last updated Oct 1, 2026
Chat apps usually charge a flat monthly fee, but developers who use AI models through an API pay by the token: the same small chunks of text that fill a context window. Prices are quoted in US dollars per million tokens, with one price for what you send in and a second, higher price for what the model writes back. Once you can read those two numbers, you can estimate the cost of almost any project on the back of an envelope.
Input and output tokens
Input tokens are everything the model reads: your question, the instructions the app adds behind the scenes, earlier messages in the conversation, and any documents you attach. Output tokens are everything it writes. For reasoning models, the hidden thinking they do before answering is usually billed as output too, even if you only see the final reply.
A price shown as $4 / $20 means $4 for every million input tokens and $20 for every million output tokens. Since one token is roughly 0.75 English words, a million tokens is about 750,000 words. Most requests are far smaller than that, so a single message typically costs somewhere between a fraction of a cent and a few cents.
Why output costs more
Look down any price list and output is almost always the more expensive side. Claude Opus 5.5 charges 5 times as much for output as for input. GPT-5.5 charges 6 times as much, and GPT-5 8 times.
The reason is how the hardware does the work. When a model reads your prompt, it can process all of those tokens in parallel, in one large and efficient pass. When it writes, it produces one token at a time, and each new token depends on every token before it. Each of those steps means reading the model's weights, plus a growing record of the conversation, out of memory again. Writing a token ties up expensive chips for much longer than reading one, and the price reflects that.
The ratio is not fixed, though. DeepSeek V4 Pro is listed at $0.435 in and $0.87 out, only twice as much, and Mixtral 8x7B is listed at $0.70 for both.
A worked example
Imagine a customer support assistant. Each request sends about 3,000 input tokens (instructions, the customer's history and their question) and gets back a 500 token reply. Here is one request on Claude Opus 5.5, at $4 in and $20 out:
- Input: 3,000 ÷ 1,000,000 × $4 = $0.012
- Output: 500 ÷ 1,000,000 × $20 = $0.010
- Total: $0.022, a little over 2 cents
Now scale it to 10,000 requests a day. That is 30 million input tokens and 5 million output tokens. The daily bill is 30 × $4 plus 5 × $20, which is $120 + $100 = $220. Output is only one seventh of the tokens, yet it is almost half of the cost.
Here is the same workload on a few other models from the catalog:
| Model | Price per 1M tokens, in / out | Daily cost |
|---|---|---|
| Claude Opus 5.5 | $4 / $20 | $120 + $100 = $220.00 |
| GPT-6 Sol | $2 / $10 | $60 + $50 = $110.00 |
| Gemini 3.8 Flash | $0.75 / $3.75 | $22.50 + $18.75 = $41.25 |
| DeepSeek V4 Flash | $0.15 / $0.60 | $4.50 + $3.00 = $7.50 |
For exactly the same traffic, the most expensive option here costs about 29 times as much as the cheapest.
Discounts: caching and batch
List prices are a starting point. Two common discounts can cut a bill a lot, but the details vary by provider, so check the current pricing page before you plan around them.
Cached input
Many requests begin with the same long block of text, such as a system prompt, a product manual or a codebase. With prompt caching, the provider keeps that repeated beginning ready and charges a much lower rate when a later request reuses it. Some providers cache automatically, others ask you to mark what should be cached, and some charge a little extra the first time the cache is written. Caches also expire after a period without use, so the savings depend on how often the same prefix comes back.
In the support example, most of those 3,000 input tokens are probably identical instructions. Caching them would shrink the input side of the bill sharply. The output side would not change at all.
Batch processing
If you do not need answers right away, for example when tagging a million product reviews overnight, many providers offer a batch mode. You submit a large set of requests and collect the results later, often within a day, at a reduced price. Discounts of around half are common, but not universal.
Other things that move the price
- Some providers charge a higher rate once a single prompt passes a certain length.
- The same open weights model can cost different amounts from different hosting companies.
- Prices fall over time. GPT-4, an early flagship, is listed at $30 in and $60 out. GPT-6 Sol is listed at $2 and $10.
Comparing cheap and flagship models
The price per token is only half the story. A few questions make for a fairer comparison:
- What does one finished task cost? A model that writes long answers, thinks for thousands of tokens, or needs a second attempt can cost more per task than a pricier model that gets it right the first time.
- How good is it at your task? Rankings like LMArena give a rough sense of quality, and cheaper does not always mean weaker. Gemini 3.8 Flash scores 1496 on the LMArena text leaderboard, ahead of GPT-5.5 at 1471, while charging $0.75 in and $3.75 out against GPT-5.5's $5 and $30.
- What is your mix of input and output? Summarizing long documents is input heavy, so the input price matters most. Drafting long reports is output heavy. Some comparison sites quote a single blended price that assumes three input tokens for every output token. For Claude Opus 5.5, that works out to (3 × $4 + $20) ÷ 4 = $8 per million.
- Do you need the extras? A larger context window, image input or stronger reasoning can justify a higher price, but only if your task uses them.
A common pattern is to send most traffic to a fast, low cost model and route only the hardest requests to a flagship.
In the quiz
Modeldle price questions ask things like "Which costs more per output token?", "Which is cheaper per input token?" or what one model costs per million output tokens. Input and output are separate prices, and the gap between them varies from model to model, so read each question closely. When the cards flip, you see the real price with a link to its source. Try today's quiz.