Comparing AI model prices: how to avoid pitfalls in price tables

Price tables on the internet look like ready comparisons, yet they lead to errors because they simplify what is actually complex. The price per million tokens is only one of several numbers that affect the bill. This text takes the price pages of three makers from 20 October 2026 and shows what to look for in them. If you have not yet calculated the cost of a task, begin with the article Task cost in tokens.

The same job at three makers

Let me compare the cost of one identical job: a thousand calls of 2000 input and 500 output tokens. I use prices of balanced and flagship class models from Anthropic, OpenAI and Google.

There are two conclusions from this comparison. First: prices within classes are surprisingly close between makers, so price alone rarely decides the choice. Second: the position of Gemini 4 Argon shows a trap discussed below, because its price changes over time. Note also that for Haiku 5.5 Anthropic gives a price with “from”, so I do not put it in the table without checking the variant.

Five pitfalls in price tables

A price with a schedule. Google’s page gives the input and output price of Gemini 4 Argon as valid until 30 November 2026, and a doubled one from 1 December. Whoever planned a budget on the first number will be surprised. A “from” price. The word means a lower bound for one variant. Thinking tokens. Google states that the output price includes thinking tokens, so reasoning increases cost regardless of the length of the visible answer. Caching and batch. Cached tokens have a separate, lower price, and batch processing a discount, which changes the cost of repeated and non-urgent tasks. The free tier. According to Google’s pricing, free-tier data may be used to improve products, while on the paid tier it is not, which matters for confidential content.

1. 01
   SourceThe official price page, not a blog roundup.
2. 02
   DateRecord the reading date.
3. 03
   ScheduleLook for change dates and promotional periods.
4. 04
   VariantLevel, mode, surcharges, 'from' notes.
5. 05
   TaskCalculate the cost of your task, not just the price.
6. 06
   ReviewRepeat at every model change.

How to read a price page

When you open an official price list, always go through it in the same order. First find the name and the ID of the model, because one family has many variants. Then read the input and output price and the unit, usually a million tokens. Next check the extras: caching, batch processing and separately priced tools, such as search with grounding, which makers bill per number of requests. Finally look for notes about deadlines and changes. Such a routine takes ten minutes and saves unpleasant surprises on the bill.

Price is not everything

The cheapest model in a table is rarely the cheapest solution. If a weaker model needs three retries instead of one, its cost per correctly done task is threefold, and on top of that come time and the user’s frustration. So compare the cost per task that passes the test, not the cost of a single call. How to choose a class and check it with a test is in the article How to choose a model for the task. For small teams predictability counts as well: a fixed subscription is sometimes better than per-token billing when traffic is irregular and the bill must not exceed a set limit.

Finally an organizational recommendation: keep in the repository a file with the prices of the models you use, the reading date and the address of the page they came from. When a maker changes its price list, you will see the difference in the change history and estimate the effect faster. More about reading model pages is in the text Frontier models: how to read vendor model pages, and the whole topic is in the Model and API pricing hub.

Last checked: 20 October 2026. Prices are a snapshot from makers’ pages; check the current ones before deciding.