Methodology
How every price on this site is verified
67 current models carry a verified price today; 56 of them were re-checked in the 7 days to October 1, 2026. This page describes what happens to each figure, from the first reading to a correction, and what the checks do not cover.
What the table covers
Models charged per token, with text output, and a price published by the maker. Audio, image, video and embedding models, and anything priced per minute, stay out even when the same page prices them. A row can enter and later be found outside that scope; it then leaves the table, the calculator and the counts, keeps its page out of the search index, and the change is recorded. Today that is Gemini Omni Flash, Gemini Omni Flash Preview: the maker's own page gives its output as video.
The first reading
Every figure is read on a page the maker publishes: its pricing page, a model page or its documentation. The row records the figure, the page it came from and the day it was read, and every page on this site shows that date and links to that page. What each kind of figure means (input, output, cache, tiers, second price lists, regions, peak hours) is explained in How LLM API pricing works. Promotional prices carry their end date and the list price they return to. Where a maker prices by region, the row names the region shown (for Alibaba, Singapore, the region it calls International) and the other level is published next to it. Prices in another currency are not converted.
"Verified" covers two outcomes: the price was read on the maker's page, or the maker was confirmed to publish no per-token price at all. 5 current rows say "no published price" for that reason, dated the day the absence was checked. No row shows an estimated, rounded or remembered number.
The weekly detector
Every Monday a script downloads the page each row cites, for OpenAI, Anthropic, Google, xAI, Mistral and DeepSeek (55 current rows today), and the pricing pages of the inference hosts Together AI, Fireworks AI, DeepInfra and Groq (20 offerings of open models today), and asks an AI model (Claude) whether the page shows the row's input and output price. The model never returns a price: it answers match, mismatch or not found, and must copy the line of the page that proves it. That quote is checked by the script: if it does not appear verbatim in the downloaded page, the answer is discarded. A quote with both figures of the row, found on the row's own page, lets the row's verification date move to the day of the check; those proposals are applied by the editor, not by the script, and a single invalid quote in a run voids them all. A mismatch, a page that changed its layout or a row no longer on its page is sent to the editor, who reads the page and corrects the row by hand; an offering a host no longer lists is reported the same way, and never removed without a person reading the page. The detector also flags a page that calls a price promotional when the row does not say so.
Alibaba (Qwen), Cohere and Meta are not covered by the detector and are checked by hand; Meta serves its price table only to a browser. OpenRouter is checked every day, not weekly: each of its 6 rows is compared with the maker's own endpoint there, and a matching price has its date moved automatically.
The Tuesday review
Some things no script sees: prices loaded by JavaScript, tabs that look the same once a page is turned into text (OpenAI's pricing page prints Standard, Batch, Flex and Fast tables with the same headers), and the absence of a price. Every Tuesday the editor opens those pages in a browser: Meta's model pages, the rows with no published price, and one tabbed pricing page, in rotation. The review is logged even when it finds nothing.
New models and maker descriptions
The detector checks the rows that exist, so a model a maker adds does not trigger it. A second weekly scan reads the makers' pricing pages without any AI model and lists every model priced per token that has no row; each one is read by hand before it enters. The same run checks that the sentence quoted on each model page, the maker's own description of the model, is still on the maker's page word for word.
Corrections and the record
A price is never overwritten silently. Every change is recorded with its date, the page it came from and an explanation, and shown on the page of each model it affects. The record separates who moved the figure: changes by a provider (4), corrections of this site's own reading (4), rows this site had missed (2) and changes of this site's criteria (7). One correction is told in full in the post about OpenAI's pricing tabs: for seven days in September 2026 this site published three OpenAI models at half their price, read from the wrong tab.
What this does not guarantee
Providers change prices without notice, and a weekly check can be up to a week behind. The authoritative source for every price is the maker's page linked next to it: confirm there before committing spend. The data is published under the CC BY 4.0 licence and can be downloaded as JSON or CSV from the dataset page. To report an error, write to the editor from the About page.