HOW THIS WORKS

How we verify every price

No scraping, no aggregator feeds, no estimates. Here's exactly what goes into a row on this site, and what we deliberately leave out.

We read the source

Every price comes from the provider's own public pricing page — never a third-party aggregator, a scraped API response, or an estimate. Where a model is open-weight (like Llama), we name the specific host whose price we're showing, since different hosts charge different rates for the same model.

We date every check

This is a new tracker — every row currently reflects a single verification pass on September 10, 2026. We re-check every price every 48 hours; any change gets a new date and gets logged on the Changelog rather than silently overwritten.

We keep it free and honest

No account needed to browse. We're not sponsored by, or affiliated with, any provider listed here — every row is sourced and treated the same way, and nobody pays for placement.

What "input" and "output" mean

Input tokens are what you send the model — your prompt, retrieved context, documents, conversation history. Output tokens are what it generates back. Output is almost always priced higher than input, often 3–5x, because generating text is more computationally expensive than reading it.

A "token" is roughly ¾ of an English word, though this varies by tokenizer and language. Providers bill per 1 million tokens; we display every rate that way for consistency, even when a provider's own page uses a different unit (per 1K tokens, for example).

What we standardize on

Most providers publish more than one rate per model — a short-context tier and a long-context tier, a standard rate and a cached-prompt rate, on-demand versus batch or provisioned throughput. To keep the comparison table apples-to-apples, we show the standard, short-context, on-demand rate for every model, and note any other published tier in that model's notes or on its dedicated page.

Some rates are explicitly promotional or time-boxed (a launch discount that steps up on a stated date). Where that's disclosed on the provider's page, we note the step-up date rather than silently treating the promotional rate as permanent.

What we don't do

We don't track benchmark scores or quality rankings — this is a price comparison, not a capability leaderboard.
We don't fabricate trend lines. Because this tracker only recently started, there's no historical price data yet; any "change since last check" figure you eventually see here will be real, dated, and sourced the same way as the base price.
We don't list every pricing tier a provider offers (batch, fine-tuning, caching, provisioned throughput) — only the standard API rate, to keep the table comparable across providers.