What we track

Which vendors, which models and which rates this site publishes — and, in the same breath, what it leaves out. Coverage is partial by design, and the shape of the gaps is stated rather than implied.

In scope

If the model you use is here, its price is here with the page it was read from. If it is not, that is a gap and not a zero.

6 providers, 147 models we publish a price for, and 734 published prices. Every tier a vendor quotes a token rate at, and every variant within it — not just the headline rate. Look your model up by name, or take the whole set as CSV or JSON.
Text and chat completion tokens, priced in USD per 1,000,000 tokens. One basis, stated as a column in the exports rather than assumed, so the first non-USD rate is visible instead of quietly wrong.
Rates a vendor has announced for a future date. An effective date is stated on a vendor's own page and never reaches an aggregator's API; upcoming prices sets each one against what is in force today.

Out of scope

Excluded at extraction time, not dropped later. A price for one of these is not something this site is quietly missing — it is something it never reads.

Everything that is not a text or chat completion token. Embeddings, image generation, audio, moderation, fine-tuning surcharges, and tool or web-search add-on fees.
Vendors and models we do not cover. Not every vendor is here, and not every model a covered vendor sells is here.
Anything announced only in prose. We read served HTML and JSON. A rate that exists only in a blog post, a slide, or a client-rendered page does not reach this site at all.

Providers: 6
Models tracked: 470including unpriced
Observations stored: 92,712append-only
Published this run: 734
Archive reaches back to: 2023-09-20

The documents

These are the normative texts, in full, with a stable anchor on every numbered section. Where a rule has a known hole, it is stated in the same paragraph as the rule.

METHODOLOGY — how a number is produced: sources and trust order, what verified and confirmed each mean, how conflicts resolve, what we will not publish, and what happens when we get one wrong. source

A disclaimer and terms of use are drafted and not yet in force, so they are not published here and nothing on this site links them as though they were. Until they clear, METHODOLOGY above is the whole of what TokenDrift commits to, and the exports are the sanctioned way to take the whole of the data.

Standing limits

Known and unresolved. The full table is in METHODOLOGY §9; these are the ones that change how you should read this site.

Some published values are explicitly unconfirmed. 174 of 734 published values could not be confirmed under METHODOLOGY §3 — the readings of the page did not agree, or only one pass saw the value. They are shown with their full provenance precisely so you can see that they are unconfirmed, and they are barred from raising a change event.
Most values are reported by only one source. 586 of 734 carry no independent corroboration at all: nobody else published that key this run, so the value is exactly as strong as the single source behind it. Only two sources ever report the same key here, so at most one can ever corroborate a price — nothing on this site says that several sources agree, because none can.
We do not warrant a maximum age for any value. verified_at tells you how old a reading is; it does not promise the reading will be refreshed within any interval.
There is no notification mechanism. A correction is labelled, never announced. If you cite a number in a context where a later correction would matter, retain what you cited and re-check it.

The observation ledger

The archive's own bookkeeping: which run is published, how many observations are stored, which reader produced them, and which source each published value came from. Useful for auditing a number, and not what most readers came for — so it opens rather than sitting in the way.

Run, readers and sources Open the ledger

What the archive holds

Published run	#82
Run finished	2026-10-17T12:01:20+00:00
Oldest observation	2023-09-20T20:39:22+00:00
Newest observation	2026-10-17T12:00:13+00:00 (just now)
Providers	6
Models	470
Observations stored	92,712
Values published	734
Of those, unconfirmed	174
Of those, with no price on either side	9
Published basis	USD per 1,000,000 tokens
Scope	text and chat completion tokens only

Observations by reader

A reader is either a deterministic parser written for one specific page, the language-model pass, or the third-party git backfill. Counts are over every observation ever stored, not just the published run.

Reader	Observations	What it is
`openrouter_json`	32,580	Deterministic parser over the OpenRouter models JSON feed. No language model reads that source.
`llm`	21,536	Language-model pass at temperature 0, read twice. Forbidden by prompt from inferring, estimating or converting units, and it cannot confirm a value the page's own parser did not emit.
`openai_astro_props`	13,200	Deterministic parser written for one specific page. It must prove its own coverage before its output is used at all.
`unknown`	7,231	Not a reader: a claim about who read the number that we withdrew. The value beside it is real and keeps its source, its URL and its verified_at — what is missing is only our ability to name which reader produced it. Never an unread or missing price. METHODOLOGY §3.
`gemini_pricing_tables+llm`	6,220	Two readers, independently, on the same page: the gemini_pricing_tables parser and a temperature-0 language-model pass both produced this value and agreed on it. METHODOLOGY §3 calls this the strongest provenance claim in the dataset.
`gemini_pricing_tables`	3,834	Deterministic parser written for one specific page. It must prove its own coverage before its output is used at all.
`litellm_git_backfill`	2,930	A public git history of a third-party pricing file. Never verified by us — it extends the series backwards to 2023 and is flagged wherever it appears.
`zhipu_pricing_table+llm`	1,512	Two readers, independently, on the same page: the zhipu_pricing_table parser and a temperature-0 language-model pass both produced this value and agreed on it. METHODOLOGY §3 calls this the strongest provenance claim in the dataset.
`anthropic_pricing_table+llm`	1,140	Two readers, independently, on the same page: the anthropic_pricing_table parser and a temperature-0 language-model pass both produced this value and agreed on it. METHODOLOGY §3 calls this the strongest provenance claim in the dataset.
`openai_astro_props+llm`	1,135	Two readers, independently, on the same page: the openai_astro_props parser and a temperature-0 language-model pass both produced this value and agreed on it. METHODOLOGY §3 calls this the strongest provenance claim in the dataset.
`xai_models_blob`	785	Deterministic parser written for one specific page. It must prove its own coverage before its output is used at all.
`deepseek_pricing_table+llm`	336	Two readers, independently, on the same page: the deepseek_pricing_table parser and a temperature-0 language-model pass both produced this value and agreed on it. METHODOLOGY §3 calls this the strongest provenance claim in the dataset.
`deepseek_pricing_table`	180	Deterministic parser written for one specific page. It must prove its own coverage before its output is used at all.
`zhipu_pricing_table`	66	Deterministic parser written for one specific page. It must prove its own coverage before its output is used at all.
`xai_models_blob+llm`	27	Two readers, independently, on the same page: the xai_models_blob parser and a temperature-0 language-model pass both produced this value and agreed on it. METHODOLOGY §3 calls this the strongest provenance claim in the dataset.

Sources behind the published run

Where the currently published values came from, and how many of them each source is responsible for.

Source	Published values	Unconfirmed	URL
OpenAI API pricing (docs)	323	76	source page
OpenRouter models API	214	—no value published	source page
Gemini API pricing (docs)	162	87	source page
Z.ai GLM pricing (docs)	22	7	source page
DeepSeek API pricing (docs)	12	4	source page
xAI models and pricing (docs)	1	—no value published	source page