How we get prices
Where prices come from
Every price on this site is a provider's list price for its own API: Anthropic's price for Claude on the Anthropic API, OpenAI's price for GPT on the OpenAI API, and so on. We collect them from models.dev, an open-source database of AI models and their prices (Copyright (c) 2025 models.dev, MIT License). Prices are in US dollars per 1 million tokens.
The same model can cost a different amount on a cloud platform (such as Amazon Bedrock or Google Vertex AI) or through a reseller. We do not list those prices yet. A model that is only available through third-party hosts is not listed.
For Chinese companies we use their international platforms, which list US dollar prices. ByteDance's Seed models (sold as Doubao on Volcengine Ark in China, priced in Chinese yuan) are not in models.dev on ByteDance's international platform, so we enter the US dollar prices from the BytePlus ModelArk pricing page ourselves. They are marked "Verified" with the date we checked them, and we recheck them at least every 45 days. See Chinese AI models.
Where we have checked a price against the provider's official pricing page ourselves, it is marked "Verified" with the date we checked it and a link to that page. Everything else is marked "models.dev". Batch prices are only shown where we have verified them, because models.dev does not list them.
A verification stops counting as soon as something may have changed: if the price at models.dev moves after the day we checked, or the official page says the current price ends on a date that has passed, the "Verified" mark is removed and the page shows the models.dev price until we check the official page again.
How often prices are checked
Every 6 hours. When a price changes, the affected pages are rebuilt and the change is added to the price changes log. We started tracking on Oct 8, 2026; the current data was checked Oct 10, 2026.
Before publishing, each update is validated. If the data cannot be read, fails validation, or the number of models drops sharply, the update is skipped and the site keeps showing the previous version.
What the labels mean
- Not published: the source does not list this price. We do not guess and we do not show it as $0.
- Free: the provider's list price is zero.
- Cache read / cache write: the price for input tokens served from, or written to, the provider's prompt cache.
- Long-context prices: some models charge more once a prompt is longer than a threshold. We show each tier.
- Batch: the price for asynchronous batch requests, where the provider offers them.
- Off-peak: some providers (such as DeepSeek) charge less outside peak hours. The main price shown is the standard (peak) price; the off-peak price and hours are listed separately.
- Deprecated: the provider has deprecated the model. We keep its page for reference.
- Beta: the provider marks the model as beta or experimental. Its price may change. (Models with "preview" in their name are shown as the provider lists them.)
How monthly costs are estimated
Cost = input tokens × input price + output tokens × output price, over 30 days. Cached input is charged at the cache read price (or the input price if no cache price is published). Cache write fees are not included. If a prompt is longer than a model's long-context threshold, that tier's prices are used. For models with off-peak pricing, estimates use the standard (peak) price. If a needed price is not published, we do not show an estimate.
Which models get comparison pages
Comparison pages are generated between 21 featured models. They are chosen by a fixed rule, not by popularity (we have no usage data): the newest general-purpose text models from major providers, one model per series, leaving out deprecated, beta and free models. Per provider: OpenAI 3, Anthropic 3, Google 3, DeepSeek 2, xAI 2, Alibaba 2, Moonshot AI 1, Z.AI 1, MiniMax 1, Mistral 1, Meta 1.
Ads and affiliate links
Some pages may show ads or affiliate links. They are labelled as sponsored and have no influence on prices, rankings or which models are listed.