Tokens are the billing unit
LLM providers bill text usage in tokens, usually priced per 1M tokens. A token can be a word fragment, word, number or punctuation mark depending on the tokenizer. Exact token counts differ by provider, so planning tools should label token counts as estimates unless they use the exact provider tokenizer.
A request normally has two billable sides: input tokens sent to the model and output tokens generated by the model.
Output tokens usually cost more
Output tokens often cost several times more than input tokens because generation is computationally expensive. Anthropic lists Claude Sonnet 4.6 at $3 input and $15 output per 1M tokens, while Claude Haiku 4.5 is listed at $1 input and $5 output. Google Gemini 3.5 Flash-Lite is listed at $0.30 input and $2.50 output per 1M tokens.
This pricing pattern means short prompts with long answers can be more expensive than expected. It also means summary detail settings matter: a detailed summary can cost noticeably more than a concise one.
| Provider example | Input price / 1M | Output price / 1M | Output multiple |
|---|---|---|---|
| Claude Sonnet 4.6 | $3.00 | $15.00 | 5x |
| Claude Haiku 4.5 | $1.00 | $5.00 | 5x |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | 8.3x |
| DeepSeek V4 Flash | $0.14 cache miss | $0.28 | 2x |
Cached input changes the equation
Some providers publish separate cached-input rates. DeepSeek, for example, lists V4 Flash cache-hit input at $0.0028 per 1M tokens and cache-miss input at $0.14 per 1M tokens. That gap can be significant for repeated prompts, system instructions or reused context.
Caching is not automatic savings for every app. The application has to send repeatable context in a way the provider can cache, and cache storage or minimums may apply depending on the provider.
Estimate using real samples
The safest budget estimate uses real examples from your product. Paste representative content into the calculator, pick an expected response type and compare per-interaction and monthly spend across providers.
Sources
Pricing changes over time. These examples use official provider pricing pages checked on July 22, 2026.
Token Pricing FAQ
Are tokens the same as words?
No. Tokens are text units used by the model tokenizer. A token can be part of a word, a whole word, punctuation or a number.
Why are output tokens more expensive?
Generating output requires model inference step by step, so providers usually price output tokens higher than input tokens.
Do cached tokens always save money?
Cached input can reduce cost when the same context is reused, but savings depend on the provider's cache rules and how your app sends prompts.