ShareYourGroupShareYourGroupAI Model Prices

How to compare AI models by cost

A price per million tokens means little until you put your own task into it.

Read on 14 November 2026. Prices and limits change: check the source.

Steps

  1. Measure a typical request. How many tokens go in (instructions, documents, history) and how many come out? The token estimator in this umbrella gives a range, and the makers' own counters give exact numbers.
  2. Use both prices. Cost = input tokens ÷ 1,000,000 × input price + output tokens ÷ 1,000,000 × output price.
  3. Multiply by requests per day or month, with a margin.
  4. Check the details of the price. Dated changes, higher prices for long prompts, separate prices for audio or images, and batch discounts all change the result.
  5. Test quality yourself. A cheaper model that gets your task wrong costs more than a dearer one that gets it right. This site gives no quality scores.

Worked examples

At the standard prices of 14 November 2026, the cost of single tasks is:

TaskClaude Sonnet 5.5Claude Haiku 4.5Gemini 2.5 Flash-LiteGemini 4 Hydrogen
Summarize a 10-page report (about 8,000 tokens in, 600 out)$0.0220$0.0110$0.0010$0.0165
Answer a customer question with a short context (1,500 in, 300 out)$0.0060$0.0030$0.0003$0.0045
Review a large code file (60,000 in, 3,000 out)$0.1500$0.0750$0.0072$0.1125

The numbers are small per request and add up at volume: multiply by your number of requests.

Things the price list does not tell you

Should I always choose the cheapest model?

No. Choose the cheapest one that passes your own tests.

How often should I recheck prices?

When you plan a budget and again when a model changes: this page was read on 14 November 2026.

Sources

  • Anthropic and Google pages named on the other pages of this section, read on 14 November 2026

Read on 14 November 2026.

This is an independent guide. It is not connected with Anthropic, Google or any company named here. Models and prices change quickly: every figure was read on 14 November 2026 on the maker's own page named under Sources. There are no quality scores here, only what the makers publish. Check the pages again before you decide.