Per-second billing
Used as published.
usd_per_second = published rate
Methodology · normalization rules · re-checked every 12h
Methodology
Four incompatible billing schemes — per-second, per-clip, per-token and credits — reduced to one comparable number: USD per second of generated video. Here is exactly how, and what we refuse to rank.
per-secondper-clipper-tokencredits✓ 87/93 verified
Re-checked every 12h · every row links to its public source · last verified Aug 1, 2026
Four billing schemes → one number
- Per-second (e.g. fal.ai, PiAPI) — used as published.
- Per-clip (e.g. WaveSpeed duration grids) — clip price ÷ clip duration.
- Per-token (e.g. Seedance on BytePlus) — (width × height × 24fps × seconds) ÷ 1024, costed at the per-token rate, so cost scales with resolution area.
- Credits (e.g. Kie.ai) — credits per generation × dollar-per-credit ÷ duration.
Everything reduces to USD per second of generated video — the one number the price pages never print. 87/93 rows verified.
Normalization · four schemes, one number
The metric: USD per second of generated video
AI video providers bill four different ways, which makes raw price pages impossible to compare. We convert everything to one number — the cost of one second of output video.
Per-clip billing
Clip price ÷ clip duration.
usd_per_second = clip price ÷ clip duration
Token billing
Tokens = (width × height × 24fps × seconds) ÷ 1024, costed at the published per-token rate for that resolution tier. Note this means cost scales with resolution area — 4K ≈ 11× the 480p price, not 4×.
tokens = (width × height × 24fps × seconds) ÷ 1024
Credit systems
Credits per generation × the provider's dollar-per-credit rate, ÷ duration.
usd_per_second = credits × $/credit ÷ duration
Every row carries a
normalization_note explaining its derivation (hover any price), plus the source URL and verification date.Comparability · what we refuse to rank
Like-for-like rules
A cheaper number is only cheaper if it buys the same thing. These four rules decide what may sit in the same ranking.
Ranking rules
- "Cheapest" rankings compare the ~720p class by default; other resolutions are listed but not ranked against it.
- Audio-inclusive and audio-extra rates are flagged per row (some providers charge ~50% more with audio).
- Fast/standard/pro tiers are labeled and never silently mixed.
- verified = we checked the provider's published price directly. reported = from a credible secondary source, pending direct verification.
Verification · 12-hour cycle
Verification pipeline
Prices are re-checked on a 12-hour cycle by automated scrapers against provider pages.
40%
Any price that moves more than 40% is quarantined for human review instead of auto-publishing. Manual overrides (for providers without scrapeable pages) carry their own verification dates.
Open data · free, no key
Free dataset, API & MCP server
The same normalized numbers the site runs on, available to your code and to your AI assistant.
Query it from your AI assistant
AI assistants can query the data directly via our MCP server (streamable HTTP, free, no key): add https://genrates.com/api/mcp to Claude Code, Cursor, or any MCP-capable client.
claude mcp add --transport http genrates https://genrates.com/api/mcp
list_modelsget_cheapestlist_pricescompare_providers
The full normalized dataset
The full normalized dataset is free at /api/v1/prices (no key required), and the cheapest-per-model summary at /api/v1/cheapest. Attribution with a link to GenRates is appreciated. The data is published as factual price information with source attribution.
93 price points · 11 providers
Changelog · 6 most recent entries
Changelog
Every dataset change is dated and typed — additions, removals, price moves and notes.
| Date | Type | Change |
|---|---|---|
| 2026-08-01 | new record | Seedance 2.5 tracked: first verified API pricing (BytePlus ModelArk, $0.103/s 480p · $0.231/s 720p t2v, native audio) — ~1.5× Seedance 2.0's rate. fal/WaveSpeed/Replicate/Kie/Novita list it as coming soon; Higgsfield advertises but access not live; Loova sells access subscription-only. |
| 2026-07-31 | removed | Data hygiene: removed 3 exact-duplicate records (same provider/mode/tier/resolution/audio + price, double-ingested during dataset merges); relabeled 1 OpenAI batch-rate record(s) to tier "batch". New automated validator now fails any build containing duplicates or contradictions. |
| 2026-07-31 | note | Deals engine live: 17 deals across 9 consumer platforms (Higgsfield, Krea, Freepik, Dreamina, Pollo, PixVerse app, OpenArt, Kling app, Hailuo app), each verified with source + date. |
| 2026-07-31 | removed | Integrity purge: removed 2 wrong-model-version proxy records (Novita "Seedance 2.0" was Seedance 1.5 Pro pricing; Kie "Kling 3.0" was Kling 2.1 pricing). Rule: a different model version is a different product — never a proxy. Caught by the owner. |
| 2026-07-30 | new record | Dataset expansion: +38 price records, +7 models (Runway Gen-4.5, Luma Ray 2/3, LTX-2, PixVerse V6, Vidu Q3, HunyuanVideo 1.5), +3 official providers (Runway, Luma, Vidu). |
| 2026-07-30 | init | Initial dataset: 58 verified price records across 6 models and 9 providers, hand-verified against provider pages. |
Corrections · 24-hour turnaround
Corrections
Wrong number, stale price, missing provider — tell us and it gets fixed the same day.
24h
Spotted a wrong or stale price? Email hello@genrates.com with a source link — corrections ship within 24 hours.