Token economics is the way AI providers charge for their language models. You pay per token, a piece of text of about three quarters of an English word, for everything you send and everything the model writes back. The bill therefore depends on how much text goes in and out, and on the model you choose, which can make a difference of a hundred times.
How is the price calculated?
Providers publish a price per million tokens, separately for input and output. In September 2026 (Anthropic, OpenAI):
- Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens, Claude Sonnet 5 $2 and $10, Claude Opus 5.5 $4 and $20.
- GPT-6 Luna costs $0.10 and $0.50, GPT-6 Sol $2 and $10, GPT-6 Astra $10 and $50.
In both families output costs five times as much as input. Writing is also the slow part of a request: a model reads the whole prompt in one pass, but writes its answer one token per step (see KV cache). The cost of a request is its input tokens times the input price plus its output tokens times the output price, divided by a million.
Where do costs pile up?
- A chat sends its whole history with every new question, and you pay for that history again each time (OpenAI). If every turn adds 500 tokens, the tenth question carries about 5,000 tokens, and the ten turns together send 27,500 input tokens.
- Reasoning models also write hidden reasoning tokens, which are billed as output (Anthropic).
- Tokenizers differ. The same text can become more tokens with one model than with another, so compare the cost of a whole task, as the token entry explains.
- Very long prompts can cost more per token. For GPT-6 Sol, OpenAI charges prompts of more than 272,000 input tokens double for input and one and a half times for output, for the whole request (model page). Anthropic keeps one rate up to a million tokens for Claude 4.6 and later models.
- Tools add input: tool definitions count as input tokens, and Anthropic charges $10 per 1,000 web searches on top.
How do you lower the bill?
- Choose the smallest model that does the task well enough, and test that on your own examples.
- Cache text that every request repeats. With prompt caching, a cached token costs a tenth of the input price or less at both providers.
- Batch work that can wait. Both providers halve the price through a batch API. Anthropic finishes most batches within an hour (batch processing); OpenAI promises results within 24 hours (batch API).
- Keep answers short. Output is the expensive part, and a maximum length per answer caps it.
- Send only the passages a question needs, for example with RAG.
Example
Suppose a webshop has 5,000 product descriptions rewritten. Each request carries 800 tokens of instructions and product data and returns a 400-token text: 4 million input and 2 million output tokens in total. On Claude Sonnet 5 at September 2026 prices that costs 4 × $2 plus 2 × $10, so $28, of which $20 is for output. The texts can wait, so through the Batch API it costs $14. Claude Haiku 4.5 in a batch would cost $7, if its texts are good enough. Caching does not help here: the instructions that repeat are shorter than the 1,024 tokens Claude Sonnet 5 needs before it caches anything.