CloudCostCompareAI model token pricing explained for non-specialists
Language models are sold by the token. If you are used to cloud servers billed per hour, token pricing works differently, and a few facts explain most of a bill.
What a token is
A token is a small chunk of text, often a word or part of a word. As a rough guide, one token is a little less than one English word, so a hundred tokens is about seventy-five words. Other languages and code can use more tokens per word. Each vendor counts with its own tokenizer, so the same text can give a slightly different count on different models.
Input and output are priced separately
Every model has at least two prices, both per million tokens.
- Input: the text you send, including instructions, documents and conversation history.
- Output: the text the model writes back.
Output tokens usually cost several times more than input tokens, because generating them takes more work. A task that reads a long document and writes a short summary is mostly an input cost. A task that writes long text is mostly an output cost.
The other prices
- Cached input. When the start of your prompt repeats, such as a long instruction block, many vendors charge a much lower rate for the repeated part. Some charge a small extra to write the cache the first time.
- Batch. Work that can wait, such as nightly processing, can be sent in a batch for a discount, with results returned later.
- Long context. A few models charge a higher rate when a request is very long.
- Reasoning tokens. Models that think before answering produce hidden tokens that are billed as output.
- Tools and extras. Search, code execution or image input can add their own charges.
Discounts do not always combine, so check each vendor’s page for how caching and batch interact.
How to estimate a month
- Describe one request: how many tokens go in and how many come out. Measure a few real examples.
- Count requests per month.
- Multiply input tokens by the input price, output tokens by the output price, and add them.
- Apply caching and batch only to the share of traffic that qualifies.
- Add a margin for growth and for longer answers than you expect.
The site’s AI cost estimator does this for common tasks, such as summarizing documents or running support chats, and shows the formula behind each figure. The discounts page lists the published cache and batch rates.
Why two models can differ by more than their price
A cheaper model is not always cheaper per task. If it needs longer prompts, more retries or produces more tokens for the same answer, the difference narrows or reverses. Compare cost per finished task on your own examples, not price per million tokens alone. Smaller models suit simple classification and extraction, and larger ones suit harder reasoning. The site’s tiers, small, mid and frontier, only group models by size and price. They are not a quality ranking.
Where the figures come from
Prices are read from public price feeds and the vendors’ own pages, updated weekly, and each row names its source and the date it was read. They are list prices, and enterprise agreements can differ. A price the sources do not publish is shown as missing, never as zero. You can read the same data through the public API and MCP server.
Build or buy
You can also run an open-weight model on rented GPUs. That moves the cost from per token to per hour, and the break-even depends on how busy you keep the GPU. See self-hosting an LLM versus using an API and GPU instances.