LLM API pricing explained: cost per million tokens

Language-model APIs are priced per token, usually quoted per million tokens, with separate prices for the tokens you send and the tokens you get back. KTwo Crust 1 Beta, for example, costs $0.04 per million input tokens and $0.15 per million output tokens. Numbers like these are easy to read and surprisingly hard to turn into a monthly bill. This guide shows how.

What a token is

Models don’t read characters or words; they read tokens, chunks of text from a fixed vocabulary built by a tokenizer. Common words are often a single token; rare words, names and code are split into several pieces.

A widely used rule of thumb for English is that a token is about four characters, or three quarters of a word, so 1,000 tokens is roughly 750 English words. It is only a rule of thumb:

• Tokenizers differ. The same text can be a different number of tokens for different models.
• Languages differ. Tokenizers trained mostly on English split other languages into more pieces. Agglutinative languages such as Hungarian, which build long words from many suffixes, often need noticeably more tokens for the same meaning.
• Code and numbers tokenize less efficiently than prose.
• Images and audio are converted into tokens too, and a single image can cost hundreds or thousands of them.

Input and output are priced separately

The input is everything you send: the system prompt, the conversation so far, any documents, tool definitions and the new message. The output is what the model writes back.

Output almost always costs more. Input tokens are processed together, in parallel. Output tokens are generated one at a time, each needing its own pass through the model, so every output token takes more compute to produce. For Crust 1 Beta the ratio is 3.75 to 1.

The formula

cost = input tokens  ÷ 1,000,000 × input price
     + output tokens ÷ 1,000,000 × output price

At Crust 1 Beta’s prices, one dollar buys 25 million input tokens or about 6.7 million output tokens. The default on our pricing estimator, 100 million input and 40 million output tokens a month, comes to 100 × $0.04 + 40 × $0.15 = $4.00 + $6.00 = $10.00.

Worked examples

Four common workloads, priced at $0.04 in and $0.15 out per million tokens:

Two things stand out. First, input dominates most real workloads: prompts and documents are long, answers are short. Second, at compact-model prices the bill is often smaller than the engineering time spent worrying about it. That changes at scale, and it changes fast when hidden multipliers creep in.

The hidden multipliers

Most surprises on an LLM bill come from tokens nobody counted.

The system prompt, on every call

A 1,000-token system prompt is sent with every request. At a million requests a month that is a billion input tokens, $40 at Crust 1 Beta’s price, before any user has typed a word.

Chat history grows faster than the chat

Chat APIs are stateless: each turn re-sends the whole conversation. Take a 500-token system prompt and a ten-turn chat in which every message and every reply is 200 tokens. Turn one sends 700 input tokens; turn ten sends 4,300. Across the ten turns that adds up to 25,000 input tokens for 4,000 tokens of actual conversation. The cost of a chat grows roughly with the square of its length.

Reasoning tokens

Some models “think” before they answer and bill those hidden reasoning tokens as output. A short answer can carry thousands of billed tokens behind it. Check whether a model reasons, and whether you can cap it.

Retries, tools and images

Failed requests that are retried, tool definitions sent with every call, and images, each worth hundreds of tokens or more, all count.

Six ways to spend less

1. Trim the system prompt. Every token saved is saved on every call.
2. Summarize or truncate history instead of re-sending whole conversations.
3. Cap the output with a sensible maximum length.
4. Pick the smallest model that does the job. Send only the requests that need a large model to one, and route the rest to a compact model.
5. Use caching and batch discounts where a provider offers them; some discount repeated prompt prefixes or requests that can wait.
6. Measure per task, not per token. A cheaper model that needs three attempts costs more than a pricier one that gets it right the first time.

Comparing prices across models

Per-million-token prices are only comparable when the tokens are. Two models with the same price can produce different bills because their tokenizers split your text differently, one writes longer answers, or one needs more retries. The honest comparison is cost per completed task: run a sample of your real workload through each model and divide the bill by the number of tasks done well.

Frequently asked questions

How many words is a million tokens?

About 750,000 English words, using the rule of thumb of three quarters of a word per token. The real figure depends on the tokenizer and the language.

Why are output tokens more expensive than input tokens?

Input tokens are processed together in parallel; output tokens are generated one at a time, each needing its own pass through the model. That makes each output token more expensive to produce.

What does $0.04 per million tokens mean?

That every million tokens you send costs four cents, so one dollar buys 25 million input tokens. For KTwo Crust 1 Beta, output costs $0.15 per million tokens.

How do I estimate my monthly LLM cost?

Multiply your monthly input tokens by the input price and your output tokens by the output price, each per million, and add the two. For Crust 1 Beta, the pricing page has an estimator that does this for you.

Can I use Crust 1 Beta at these prices?

The prices apply to approved private-beta accounts only. KTwo currently offers no public services, and a listed price is not an offer or a promise of access. You can request access.