The same model can cost 20x more depending on the channel you buy it from. One example from today:
DeepSeek V4.1 Flash: 66 metered channels, 0.2 to 4.02 CNY per 1M tokens
Metered input prices only, normalized to CNY per 1M tokens. Subscription plans are excluded: prorating a monthly plan into a per-token number and comparing it to metered pricing produces multipliers that do not survive a second question.
What the site does
llmrankhub.com puts 630 models and 3543 channel listings in one table.
Cheapest channel. Metered and subscription pricing are labeled separately, filtered by uptime and context length, and every listing keeps its source link so you can check it yourself.
Free channels. 103 listings carry a free quota, most with daily or per-minute caps. Validate a pipeline on a free tier before you commit a card.
Unused quota. Plans expire and projects end with balance left over. That quota can be listed and resold with balance verification, stated expiry, and escrow released only after a test call succeeds.
Why the table matters
Plenty of sites show a price. The difference here is the unit and the split: everything is CNY per 1M tokens, subscription and metered never mixed, and uptime sits next to the price. Buying the cheapest listing without checking those three is how people end up paying more.
Data sources and refresh times are on the page. Prices are the channel's, at the moment you order.
llmrankhub.com
I built this site. If a price looks wrong, there is a correction form and I read it.
Top comments (1)
Choosing one unit - CNY per 1M tokens - and separating subscription from metered is the decision that makes a table usable, and it is the one most comparison sites get wrong. Two tests I would add to the pipeline, both of which catch errors nobody will ever submit through the correction form.
The first is a distinct-price ratio per category. Count the priced rows and count the distinct prices: a category with 11 priced rows and 2 distinct values is broken, and you can see that without knowing anything about the market. For reference, in the feed I audited this week the healthy case was 47 rows and 41 distinct prices, and the broken case was 11 rows and 2. That ratio runs on every refresh, costs nothing, and flags exactly the class of error a reader cannot detect, because every cell still looks normal.
The second is a same-provider constant test. If every product from one provider carries the identical price, that is a fallback value rather than thirteen coincidences. In that same feed one provider was 6 for 6 at the same number and another was 3 for 3, and both were placeholders. A rule of "one provider with more than two products and one distinct price gets marked unverified" would have caught all of it automatically, and it generalises to providers you have not audited yet.
The third is a rendering rule rather than a test: any row whose unit is not a billing unit - "flat", "variable", "custom" - has to be excluded from the ranking, not shown with a caveat. A ranked table implies the numbers are comparable, and one bad unit inside the sort order is worse than one missing row.
The correction form is the right mechanism and more than most sites offer. It just depends on a reader noticing, and the errors that cost the most are the ones that look exactly like data.