Methodology: How We Collect and Compare AI API Prices

This page explains where the numbers come from, how they are normalized and compared, and what they do not tell you.

1. How prices are collected

We read each provider’s public pricing page, model pages, docs or public pricing API. Every price row records the source URL and a confidence level: “Pricing page” (seen on the provider’s own pricing or docs page) or “Pricing API” (read from a public API response). We only record what the page shows; unknown values stay blank. We stop at login walls, CAPTCHAs and bot challenges, and we do not sign up, create API keys or buy anything to see a price.

Providers fall into three groups: official (the model's own developer), aggregators (an inference platform that hosts many vendors' models under its own brand) and resellers (sells access to other vendors' models, usually below list price). The current data set has 1,606 price rows from 50 providers, all checked on 2026-09-28. 30 rows that repeated the same provider, spec and price (recorded in two research batches) were merged.

2. Per-second and per-image normalization

• Video is normalized to US dollars per second of output. Per-second prices are used as listed. Per-video prices are divided by the clip length when the page states it; if it does not, the row is shown without a normalized figure rather than guessed.
• Images are normalized to US dollars per image. Per-megapixel prices are converted at the size recorded with the row (usually about 1 megapixel, 1024×1024, or the provider’s own example size).
• Credits are converted only at the rate the provider itself states; the note on each row keeps the original figures.
• Token-billed models (for example Seedance on BytePlus) use the provider’s own per-video examples. Where no per-unit figure is published, the row is listed without a price (3 rows).
• China (CNY) list prices (68 rows) are shown in yuan in a separate table and are not converted, because the pages state no exchange rate. We show the ratio of the CNY price to the international official USD price for the same spec.

3. Comparing with the official price: exact vs relaxed

“× official” divides a normalized price by the cheapest official price for the same model, resolution and (for images) mode. Text-to-video and image-to-video are treated as one spec because they are almost always priced the same.

• Exact: tier, resolution and audio also match.
• Relaxed (shown with ~): tier, resolution or audio is unspecified on one side, and every official spec it could correspond to costs the same.
• No comparison (—): the match is ambiguous, for example when a reseller does not say whether audio is included and Google prices the two differently. We do not force a comparison between different specs.
• Where a provider sells a separately named cheaper variant (such as Fast or Mini) that the developer does not list officially, we do not compare it with the standard model’s price.

The flagship tables on the home page use an explicit spec for each row (tier, resolution, audio and, where needed, clip length). They leave out separately named variants (Fast, Mini, Developer, Beta or relaxed-queue tiers) and bring-your-own-account estimates, which remain visible on each model page.

4. Published price is not a tested price

Every figure on this site is what a provider publishes. It does not tell you whether requests succeed, how long they take, whether failed jobs are charged, or whether the output comes from the model named. Resellers do not always disclose how they source capacity. Until we test, we label every price table “Published prices — not yet independently tested”.

5. Test generations (planned, not yet running)

We plan to run daily test generations against a sample of providers for each flagship model, recording success or failure, latency, the amount charged and the output. Test results will be shown separately from published prices, with the date and prompt used. No tests have been run yet, and no page on this site claims otherwise.

6. Update cadence

All prices were checked on 2026-09-28, and every page shows that date. The plan is a daily automated refresh for the 19 providers that publish a machine-readable price list without an API key (302.AI, AI/ML API, APIYI, Atlas Cloud, CometAPI, Crazyrouter, DeepInfra, Defapi, Eachlabs, EvoLink, Kie.ai, LaoZhang API, Novita AI, OpenRouter, PiAPI, Runware, Segmind, WaveSpeedAI and fal.ai), and a weekly check of the remaining pages. Each page will keep showing the date its prices were last checked.

7. Editorial independence

• Sponsorships and advertising never change our data, rankings or test results.
• Providers cannot pay to be included, removed, re-ordered or described differently.
• Any sponsored placement will be labeled “Sponsored” and kept separate from price tables and test results.
• This site currently contains no affiliate links. If that changes, affiliate links will be labeled and will not affect data, ordering or test results.
• We describe providers with sourced, dated facts. Reported issues are attributed to their source and are not our own findings.

8. Corrections

If a price is wrong or out of date, email hello@aivideoapis.com with a link to the source. Providers can request corrections the same way; corrections are based on the public source, not on who asks.

9. Privacy

We measure visits with Cloudflare Web Analytics, which does not use cookies or fingerprinting and does not track individual visitors across sites. We don't run ads or third-party trackers.