How to tell a real AI model price from a placeholder (26 of 116 are the same number) An engineer auditing a public AI model pricing feed of 116 models found that 26 rows share an identical price of 1.54 and 48 rows (41.4%) carry the unit 'flat' — both artifacts of scraper fallback defaults rather than real data. The defaults cluster by provider and category: every Kuaishou Kling, PixVerse and Vidu Shengshu model is priced at 1.54, and the Audio category has 11 priced models with only 2 distinct prices, while Chat/LLM shows 47 priced models with 41 distinct prices. The engineer estimates only about 21 rows are actually usable, and attributes both defects to two lines of default-handling in the ingestion path. A price list looks like a column of facts. In the feed I audited for this post, it is fact on most rows and a default value on some of them — and the defaults look exactly like the facts. That is the failure mode worth writing about, because an empty cell makes you suspicious and a filled-in default does not. 116 models. 26 of them carry the identical price: 1.54 . Twenty-six coincidences would be a story about the market. This is not that, because the duplication is not spread randomly across the table — it covers every single product of several providers : | Provider | Models in the feed | Priced at exactly 1.54 | |---|---|---| | Kuaishou Kling | 6 | 6 all | | PixVerse | 3 | 3 all | | Vidu Shengshu | 2 | 2 all | | MiniMax | 10 | 8 | | Tencent | many | 2 | When a provider's entire product line lands on one number, the most economical explanation is that the number is a fallback: the scraper could not find a price, so the field got a default. 1.54 is that default. Repeated values happen for real reasons. The test I would apply is the distribution across categories and providers , not the count. The cleanest single statistic: the Audio category has 11 models with a numeric price and only 2 distinct prices. A distribution like that cannot come from real pricing — no vendor prices eleven products across two values. Meanwhile Chat/LLM has 47 priced models and 41 distinct prices , which is what real data looks like. | Category | Models with a price | Distinct prices | Verdict | |---|---|---|---| | Chat / LLM | 47 | 41 | usable | | 3D | 9 | 6 | mostly usable | | Video | 21 | few | 14 of 22 are 1.54 | | Audio | 11 | 2 | 9 of 11 are 1.54 | The same feed has a second, independent default problem in the unit field. 48 of 116 rows 41.4% have the unit flat — which is not a billing unit. And its distribution is diagnostic: | Category | flat share | Real billing unit for that category | |---|---|---| | Chat / LLM | 71% | per token | | Audio | 73% | per token / per second | | Embedding | 67% | per token | | Video | 5% | per second / per token / per generation | | Image | 0% | per image — per generation is correct here | | 3D | 0% | per generation — 9 of 9 correct | flat fires in exactly the categories whose real unit is per-token, and never in the categories where per-generation is genuinely right. That is not noise either — it is the same pattern: when the source page does not state a unit, the field falls back to a default, and the pages that do not state a unit are precisely the per-token ones. Two fields, two defaults, one mechanism: unit missing → flat . It looks like a unit. price usd missing → 1.54 . It looks like a price. The real failure mode of a data feed is not emptiness. Empty is loud. The failure mode is a field that is populated with a plausible default, because then the table is complete and wrong in the same glance. Publishing 26 defaults as 26 real prices means a 116-row table lies on 22% of its cells — and every one of those cells looks normal. Cross-referencing both defects, the rows you can act on today are: per generation , 6 distinct prices. This is the cleanest column in the table. per 1M tokens plus the 8 labelled variable , custom , contact us are not units and cannot be compared. An index is worth something only if the reader can tell which rows to trust. If I told you this feed had 116 prices, you would reasonably assume 116 prices. The honest number is closer to 21 rows you can act on , and saying so is what makes the other rows worth reading. Notably, both defects are one rule each , not 48 and 26 individual corrections. The flat fallback and the 1.54 fallback are two lines of default-handling in the ingestion path. Data quality work is usually described as a checklist; this one is arithmetic on the defaults. All figures from the public feed at generated at 2026-11-20 19:00 UTC , 116 rows, platform markup pct: 0 . The 1.54 conclusion rests on whole-provider coverage, not on the value itself. No de-duplication or estimation was applied to any number in this post.