1. Tokens: The Unit You Pay For
Language models read and write text in tokens. For ordinary English, one token is about four characters, or roughly three quarters of a word, so 1 million tokens is around 750,000 words. Code, other languages and unusual words use more tokens per word. Every provider counts tokens with its own tokenizer, so the same text can be a slightly different number of tokens on different models.
Prices are quoted per 1 million tokens (often written “per MTok”). A price of $2 per 1M input tokens means sending 1,000 tokens costs a tenth of a cent or less.
2. Input and Output Have Different Prices
What you send (your instructions, the conversation so far, any documents) is billed at the input price. What the model writes back is billed at the output price, which is almost always higher: on our board the median output price is 5.0 times the median input price. Reasoning models also bill their hidden “thinking” as output tokens, so a short answer can still use many output tokens.
A worked example from our board: 1,000 chatbot replies of about 500 input and 300 output tokens each cost $0.16 on Ministral 3 14B, the cheapest model we track by blended price, and $20 on GPT-6 Astra, the most expensive. The calculator does this sum for your own usage.
3. Long-Prompt Tiers
Some providers charge more once a single prompt passes a size limit. Two examples from the official pages we read:
- Anthropic lists Claude Haiku 5.5 at $0.10 per 1M input tokens for prompts up to 100,000 tokens and $0.50 for prompts over 100,000 tokens; output goes from $0.50 to $2.50 per 1M at the same line.
- Google lists Gemini 3.1 Pro (preview) at $2.00 per 1M input tokens for prompts up to 200,000 tokens and $4.00 above that, with output at $12.00 and $18.00.
If your app sends long documents or long chat histories, check whether you cross a tier: the whole prompt is usually billed at the higher rate once it does.
4. Cached Input
When you send the same long beginning again and again (a system prompt, a policy document, a code base), many providers can serve it from a cache at a lower price. On our board, 38 of 42 models list a cached-input price, and the median saving is 90% against normal input. Some providers also charge a one-time fee to write to the cache; OpenAI, for example, lists separate prices for 30-minute and 12-hour cache writes. Caching pays off when the repeated part is large and is reused many times within the cache’s lifetime.
5. Batch Discounts
If you don’t need the answer right away, batch processing is cheaper. Anthropic’s Batch API gives a 50 percent discount on both input and output, Google lists its Batch API as a 50% cost reduction, and OpenAI’s batch price table shows half the standard prices (for example $4.00 instead of $8.00 per 1M input tokens for gpt-6.2-astra). Batches typically finish within hours rather than seconds, which suits overnight jobs such as tagging, translation or evaluation.
6. How to Estimate Your Bill
- Count a typical request: tokens in (instructions + context + user text) and tokens out.
- Multiply by requests per month to get monthly input and output tokens.
- Multiply by the input and output prices, and check whether a long-prompt tier applies.
- Take off what caching and batch processing can save for the parts that qualify.
- Compare a cheaper model in the same family: on our board the price gap inside one provider’s line-up is often ten times or more.
Our price board lists every price with the date it was read on the official page and keeps the full history, so you can see when a provider cut or raised a price.