What is a token? How AI pricing really works

A token is the unit AI models read and write, roughly a short word or a piece of one. Providers bill per million tokens, split into input (what you send), output (what the model writes back) and, for reasoning models, hidden reasoning tokens that are billed as output. The same task can cost a fraction of a cent or several cents depending on how many tokens it uses and what each one costs.

What a token is

A tokenizer splits text into chunks before the model sees it. Common words are usually one token. Long or unusual words, code, numbers and non-English text often take several. The split depends on the model's tokenizer, so the same text can have different token counts on different models.

Pricing pages list rates per 1,000,000 tokens. Examples from my price sheet as of 2026-10-11:

Cached input is a lower rate for input the provider has seen recently, such as a long system prompt that repeats across calls. It matters most for agents and chat apps that resend the same context again and again.

Input vs output

In the table above, output costs five times the input rate for GPT-6 Luna, Claude Sonnet 5.5, Claude Opus 5.5 and Claude Fable 5.1. The length of the answer often matters more than the length of the question.

You can check this by hand. For a single chat answer on Claude Sonnet 5.5, I measured 61 input tokens and 1,256 output tokens. At list prices:

• Input: 61 tokens at $2.00 per 1,000,000 tokens, a tiny fraction of a cent
• Output: 1,256 tokens at $10.00 per 1,000,000 tokens, almost the whole bill
• Total: 1.3 cents per task, or $12.68 per 1,000 tasks

Output is about 99 percent of that cost. A shorter answer, or a model that writes fewer tokens for the same job, changes the bill far more than a shorter prompt would.

Reasoning tokens

Reasoning models think before they answer. That thinking is generated as tokens, and those tokens are billed as output even though the user never sees them.

Take translating 1,000 words on DeepSeek V4.1 Flash. I measured 1,417 visible output tokens and 10,291 reasoning tokens, more than seven times the visible output. At $1.20 per million output tokens, the job costs 1.4 cents, and most of that is thinking the user never reads.

Reasoning also varies by model in ways the price sheet does not show. Some models have reasoning always on, some by default, some not at all. Claude Sonnet 5.5 reported 0 reasoning tokens in my runs, while GPT-6.1 Sol reported 103 on the chat task and 26 on the PDF summary. Compare the reasoning column, not just the list price, before choosing a model for a high-volume job.

Tokenizers differ

Token counts are not comparable across providers. This is the number of tokens each provider's tokenizer needs for the same English document, with OpenAI set to 1.00:

• Anthropic: 1.39
• OpenAI: 1.00
• Google: 0.98
• MiniMax: 0.99
• Mistral: 1.03
• xAI: 1.07

Anthropic's tokenizer produces about 39 percent more tokens for the same English text than OpenAI's. My PDF run shows it. The 20-page document (10,023 words of J. S. Mill's On Liberty) came to 12,987 input tokens on GPT-6.1 Sol and 18,004 on Claude Sonnet 5.5, a ratio of about 1.39 on the same content.

A comparison based only on per-million rates can mislead. A model with a lower rate can still cost more per task if its tokenizer or its reasoning inflates the count.

What tasks really cost

I ran each task on each model and recorded token counts and cost. The median across all 326 models is 0.24 cents for a chat answer, 0.83 cents for summarizing a 20-page PDF and 0.79 cents for translating 1,000 words. Twelve models are measured with real runs (one run per task, default settings, output capped at 16,000 tokens); the rest are estimated. Here is the chat task:

#	Model	Cost per task		Text score	Tokens in / out	Basis
1	GLM 5.3 FlashZ.ai	0.04¢$0.36 / 1,000		1474	48 / 1,782	measured best value
2	GPT-6 LunaOpenAI	0.06¢$0.59 / 1,000		1443	45 / 1,179	measured
3	Claude Haiku 5.5Anthropic	0.07¢$0.66 / 1,000		-	61 / 1,299	measured
4	Qwen3.8 FlashAlibaba Qwen	0.08¢$0.76 / 1,000		-	88 / 1,598	measured
5	Mistral Small 4Mistral	0.08¢$0.83 / 1,000		-	40 / 1,377	estimated
6	MiMo-V2.6-ProXiaomi	0.09¢$0.90 / 1,000		1479	206 / 933	estimated best value
7	DeepSeek V4.1 FlashDeepSeek	0.09¢$0.91 / 1,000		1475	65 / 6,320	measured
8	Llama 4 MaverickMeta	0.11¢$1.06 / 1,000		1327	87 / 1,592	estimated

The spread is wide. GPT-6 Luna's chat answer costs 0.06 cents, using 45 input, 540 output and 639 reasoning tokens. Claude Fable 5.1 costs 6.3 cents (an estimate), using 61 input and 1,256 output tokens with no reasoning. Token counts differ, but most of the gap comes from price per token.

For the PDF task, input dominates because the document is long:

Claude Sonnet 5.5 and Claude Fable 5.1 used identical token counts, so the gap between them comes from price alone. See the PDF summary task page for every model.

How to estimate your own workload

Count four things per task:

• Input tokens, including any system prompt and retrieved context
• Visible output tokens
• Reasoning tokens, if the model reasons
• The model's input, output and cached input rates

Multiply each by its rate, add them, then multiply by your monthly task count.

Agents need extra care. My agent scenario uses an 8,000-token system prompt, 50 steps, 1,500 added context tokens per step and 300 visible output tokens per step, so input grows with every step. Caching and step count matter more there than the per-token price.

The calculator runs this math. The methodology page explains how I ran each task, and the chat answer task page shows per-model results.

FAQ

Is a token the same as a word?

No. A token is often shorter than a word for common text, and a word can take several tokens in code, numbers and many non-English languages. Counts also differ by tokenizer.

Why does my bill have more tokens than my prompt?

Your bill counts the whole exchange: system prompt, retrieved context, visible output and, for reasoning models, hidden reasoning tokens billed as output. The prompt you type is often the smallest part.

Which number should I compare between models?

Compare cost per task on the same job, not list price alone. A lower per-token price can still cost more if the model writes many reasoning tokens or uses a tokenizer that produces more tokens for the same text.