AI video model free-tier data

Every dated figure published on this site is also available here as a plain dataset, so it can be checked, quoted or built on without scraping the articles. 125 observations across twelve models — Seedance 2.5 (26 rows), PixVerse (9 rows), Runway (10 rows), Wan 2.5 (16 rows), Kling AI (12 rows), Sora 2 (3 rows), Grok Imagine (5 rows), Invideo AI (7 rows), Hailuo AI (3 rows), Luma Dream Machine (16 rows), Canva (12 rows), Veo 3.1 (6 rows) — each one carrying the channel it came from, the date it was observed, and how it was obtained. A row can record a number, a state, or a documented absence — what an official page does not state is itself checkable, so it gets a row instead of an empty cell.

What is in it

The dataset records four kinds of row: what a channel grants, what one generation costs, what the vendor refuses to state, and what a visitor sees without an account.

• Free-tier grants — daily credit allowances, daily generation caps, sign-up bonuses, launch promotions, one-time free-plan credit deposits and their expiry (Runway: 125 one-time credits that do not expire), what a deposit converts into on the vendor's own line (Runway: "125 credits = 25s of Gen-4 Turbo, or 25 image generations"), which models a free plan may use at no credit cost (Runway: four models marked FREE in the model picker, 2026-10-04), and a grant stated only in machine-readable data with no amount attached (Luma Dream Machine: "All plans include free trial credits and commercial use", 2026-10-05 — and whose size is reported at about 30 generations a month from a logged-in account on the same date, by a different grade of evidence), by channel.
• Credit economics — what one generation consumes, by model and unit (Runway: 12 credits per second of generated video for Gen-4.5, stated by the vendor, and 5 per second for Gen-4 Turbo, our arithmetic on the page's own equivalence lines — so the same 125 free credits buy 25 seconds on one model and about 10 on the other, and the page's per-model cost table runs from 60 credits per 5-second generation up to 160 credits per 4 seconds, which is one Seedance 2.0 Fast generation affordable on the free grant and not a single Seedance 2.0 Pro one; Wan: 6 credits for 2 s, 15 for 5 s and 30 for 10 s, measured on a logged-in free account, which works out to 3 credits per second — our arithmetic on the measurements, not a rate the vendor states. Wan's credit packs separately imply 5 credits per video, which does not reconcile with the measured 3 per second; both figures are kept, neither is averaged).
• Hard limits — maximum clip length, extension ceiling, output resolution, free-plan storage caps (Runway: 5GB), and what the vendor actually states versus what hosting platforms advertise.
• Watermark behaviour — whether the free tier marks output, and the toggle that paid tiers still have to switch on. (Seedance 2.5 rows only, so far.)
• API reality — availability status, error codes seen in practice, and how failures are billed. (Seedance 2.5 rows only, so far.)
• Documented absences — free-tier facts that official pages do not state, recorded with the date they were read. Fourteen rows so far: four Wan 2.5 rows (2026-10-03 — no credits-per-month figure for the free plan, and no watermark-free, resolution or duration row in the free column), two PixVerse rows (2026-10-02 — the first model in this dataset whose entire free-tier record is an absence), two Grok Imagine rows (2026-10-04 — the official pages are not reachable from this environment, and the free-tier allowance carries no published figure), two Luma Dream Machine rows (2026-10-05 — the pricing page's plan table carries no free row at all, and no Luma page publishes an amount for the free tier), one Kling AI row (2026-10-03 — the public pricing URL redirects to the marketing site, and the plan table renders only inside a logged-in session), one Sora 2 row (2026-10-04 — the free-tier allowance is not measurable, because the product was discontinued) one Seedance 2.5 row (output resolution, 2026-07-31) and one Invideo AI row (2026-10-05 — the pricing page names a Free plan in its FAQ and states no figure for it). A row whose value carries content plus a disclosed gap — the extension ceiling with "no round limit stated", Wan's free-plan credit source with no amount stated, and Luma's free trial credits stated to exist with no amount — stays under its own heading, with the gap in its note.

Files

All three files are the full dataset — every model currently covered (Seedance 2.5, PixVerse, Runway, Wan 2.5, Kling AI, Sora 2, Grok Imagine, Invideo AI, Hailuo AI, Luma Dream Machine, Canva and Veo 3.1) — not Seedance-only extracts. The file names keep the seedance-2-5-free-tiers stem for link stability; the coverage lives in the content.

Fields

Reading it correctly

Two things will trip you up if you skim it.

Some rows contradict each other on purpose. Where two observations disagree — the daily credit figure in July versus August, the vendor saying the API is "coming soon" while a developer reports it opened two weeks earlier — both rows are kept. Neither was averaged into a number nobody published.

Free-tier numbers move without announcement. A row is a record of what was true on its date, not a promise about today. Check the observed column before quoting anything as current.

What the three source grades mean

Every row carries one of three grades in its source_type column, and the grade tells you what kind of checking the row has had. At this writing the dataset holds 68 vendor entries, 56 report entries and 1 single-source entry.

vendor means the statement was read off the vendor's own material — a pricing page, official documentation, or machine-readable page data — and it also covers documented absences: when an official page simply does not state a figure, that silence is recorded as a vendor-graded row with the date it was read, because the absence was verified, not assumed. The Seedance 2.5 output-resolution row ("not stated", ByteDance official pages, 2026-07-31) is graded vendor for exactly that reason.

report means the figure came from a person using the product — most often this site's own logged-in observations of a free account. Kling's free trial package (66 credits, expiring 2026-11-03, read from the logged-in banner on 2026-10-03) is a report row: the number is dated and specific, but you cannot re-read it without an account on that channel.

single-source means one account stands alone and nothing else corroborates it. The dataset's only such entry says Seedance 2.5 output resolution reached 4K + 10-bit, from one pre-launch tester under NDA, observed 2026-07-04 — the note itself states that nothing else corroborates it. Such a row is kept because it is dated and attributed, not because it is confirmed; it needs a second source before it should be treated as fact.

How to check a row yourself

Each row carries three things a checker needs: the channel it was observed on, the date it was observed, and the grade above. A vendor row can be re-read at the vendor's own page on or after its date — open the official pricing page or documentation named in the channel column and look for the same figure. A report row requires a logged-in account on that channel; some of these have hard gates — Kling's public pricing URL redirects to the marketing site and the plan table renders only inside a logged-in session, Wan's generate page presents a login modal to anonymous visitors, and Invideo's signup flow redirects — so a report row may be impossible to re-verify without registering.

A documented absence is checked the same way as any vendor row: re-read the same official page and confirm the figure is still not there. One asymmetry to keep in mind: re-reading a row later can confirm it or date a change, but it can never re-create the state of the page on the original date — so when a re-read disagrees with an old row, both are kept rather than overwritten, which is what the "Reading it correctly" section above describes.

Changes the dataset has actually caught

Because every entry carries its observed date, the dataset doubles as a change record — but only where a channel has been read more than once. At this writing exactly one channel-and-metric pair has been observed twice: the daily free credits on Jimeng, reported at 60–100 on 2026-07-04 and at 88 on 2026-08-11. Both readings stand in the table, and no announcement between those dates is recorded here, so the dataset cannot say whether the grant moved or the first report was a range that later settled. Every other figure in the dataset has been observed once, which means there is no other dated change to show — not because nothing changed, but because nothing has been re-read. An expiry date such as Kling's free trial package expiring 2026-11-03 (report, 2026-10-03) is an attribute the grant carried from the day it was recorded, not a change; a change, in this dataset's sense, is a second reading of the same channel and metric on a later date. What one twice-read figure does and does not imply about how often allowances move is discussed on the free AI video generator page.

How often this dataset updates

There is no publication schedule. Entries are added when a channel is measured for the first time or re-read, and the dates in the observed column — running from July 2026 to 2026-10-08 at this writing — are the only record of when the dataset grew. No entry is back-dated or silently overwritten; a re-read that disagrees with an older entry is added alongside it, as described under "Reading it correctly" above. If a figure here looks out of date or wrong, the way to report it is given in the Corrections section at the end of this page.

Preview

Cite it

The dataset is released under CC BY 4.0. If you use it, credit VideoFreeTier and link to videofreetier.com. There is no attribution string to copy beyond that.

Corrections

If a figure is out of date or wrong, Corrections are welcome at contact@videofreetier.com.