Every price here is dated. Not the page, the price.

What digital services cost, the limits they impose, and above all what changed, when, and with the evidence. Every figure below carries the day we read it, the page we read it on, and a fingerprint of that page.

A language model is trained on a snapshot of the world. Anything whose correct answer changes faster than the training cycle stays out of its reliable reach: prices, quotas, rate limits, end-of-life dates. It has to consult a source, and cite it. This is that source.

Services published1 297from 22 sellers

Published observations72 983each with a source fingerprint

Held back17 465parameters lacking a high enough source

Prices that are already set for a date that has not arrived

55 prices published by Google take effect on January 1, 2027. We read them on the vendor's own price page and kept a fingerprint of that page, so the announcement stays checkable after the page changes. For example, Gemini 3.6 Flash costs $3.75 per million output tokens today and $7.50 from January 1, 2027.

See every announced price

Which is cheaper?

Computed on 18 September 2026 at 15:01 UTC from prices we read at the source, each dated. At the entry level of each family, on output tokens:

Every family, compared

What you will not find anywhere else

The same model does not cost the same everywhere. Because we collect from several sellers independently, each price carrying its own observation date, we can put them side by side. No single provider's page shows this, and no language model can answer it: it requires having read all of them.

See the same model, seller by seller

Price is half the question. The other half is speed.

The cheapest offer is not a bargain if it is slow, and it is worth nothing if it is down. We publish time to first token and output throughput as the Hugging Face router measures them, and availability as OpenRouter's router measures it. Neither figure is ours, both say so, and neither is what you would get calling a host yourself: the network path is not the same.

We never show the latest reading on its own. One latency measurement covers a handful of calls. We read once a day, keep every reading, and show the median of ours with how many it rests on. The firms that measure speed cap their public history at ninety days or at seventy-two hours. We keep ours. Ask us in 2030 what a host was doing in 2026, and we will still have the answer.

What we refuse to publish: an average of what users report on forums, because an average of anecdotes is a figure nobody measured.

Speed and availability, and what our figures are worth

Why most of what we hold is not published

Sources are ranked from 1 to 5. A level 5 source, meaning a third-party aggregator, may be used to corroborate a figure or to trigger a check. It is never published on its own. Today that leaves 17 465 parameters held back, which is 55% of what we hold. The number is uncomfortable and it is honest. It will fall as we add first-party sources.

How we rank and verify sources

Recent changes

Detected	Service	Parameter	Change
18 September 2026	tencent/hy4-preview (on Novita AI)	Context length	-4.6%
18 September 2026	DeepSeek: DeepSeek Flash Latest (sold by OpenRouter)	Cache read price	-70.3%
18 September 2026	DeepSeek: DeepSeek Flash Latest (sold by OpenRouter)	Input price	-1.0%
18 September 2026	DeepSeek: DeepSeek Flash Latest (sold by OpenRouter)	Output price	-1.0%
18 September 2026	zai-org/GLM-5.2 (on DeepInfra)	Discount rate	-28.6%
18 September 2026	DeepSeek: DeepSeek V4 Flash Vision Exp (sold by OpenRouter)	Maximum output	-72.2%
18 September 2026	DeepSeek: DeepSeek V4 Flash Vision Exp (sold by OpenRouter)	Cache read price	-2.0%
18 September 2026	DeepSeek: DeepSeek V4 Flash Vision Exp (sold by OpenRouter)	Input price	-2.0%

Recently listed models

Listing dates as published by the catalogue we read them from, not the date we first saw them.

Browse the catalogue