A dated record of what each channel actually grants on the free tier of AI video models — daily credits, daily generation caps, sign-up bonuses, clip length, extension ceiling, output resolution, watermark behaviour, and API availability.
Dataset Card
By Zhou, published under cc-by-4.0, revision 509fb8392385.
AI video model free-tier observations
A dated record of what each channel actually grants on the free tier of AI video models — daily credits, daily generation caps, sign-up bonuses, clip length, extension ceiling, output resolution, watermark behaviour, and API availability.
Free-tier terms are the least documented part of an AI video product. Vendors publish a launch post with a headline number, then adjust the daily grant, move the watermark switch, or gate a model behind a membership tier. This dataset collects those figures in one place and stamps every row with the channel it came from and the date it was observed.
66 observations across Seedance 2.5, PixVerse, Runway, Wan 2.5, Kling AI, Sora 2 and Grok Imagine, current to 2026-10-04. File names keep the seedance-2-5-free-tiers stem for link stability; the files themselves are the full dataset, not Seedance-only extracts.
Files
| File | Format | Use |
|---|---|---|
seedance-2-5-free-tiers.csv |
CSV, UTF-8 | spreadsheets, pandas, csv module |
seedance-2-5-free-tiers.json |
JSON | includes field notes and the licence in the envelope |
seedance-2-5-free-tiers.md |
Markdown table | pasting into a document, an issue, or a pull request |
Fields
| Field | Meaning |
|---|---|
model |
the model the figure applies to |
channel |
the product, platform or API the figure was observed on |
metric |
what is being measured |
value |
the figure, quoted as observed — ranges and hedges are kept, not flattened |
observed |
the date the figure was read or recorded |
source_type |
vendor = stated by the vendor · report = reported by a user · single-source = one uncorroborated account |
note |
caveats, conflicts, and what is known to be missing |
Reading it correctly
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. If you need one value, decide which source you trust; the dataset will not decide for you.
These numbers move without announcement. A row is a record of what was true on its date, not a promise about today. Check observed before quoting anything as current.
Nothing here is estimated. Where a figure could not be confirmed against a primary source, the row says so in note and source_type reflects it. There are no interpolated values.
Measured, not only quoted. Most vendor rows are what a company published. Twelve rows are what happened when someone ran the product: an unauthenticated attempt to open the generator and the login wall it hit; a daily check-in grant of 6 credits that appears on no vendor page; the credits deducted by a 2-second, a 5-second and a 10-second generation; the resulting per-second rate; the free plan's duration ceiling; PixVerse's 50-credit deduction for a 5-second generation; and Kling's 66-credit trial package, its 2026-11-03 expiry and its 720P rate; and Runway's model picker, where four models are marked FREE. Both grades are marked, and the second is the rarer one.
What is deliberately not in here
- No invented price per generation. Where a vendor's own published numbers allow a rate to be derived — Wan's credit listings all work out to 5 credits per video and 0.25 per image — the derived value is recorded in that row's
noteand marked as our arithmetic on the vendor's wording, not as a unit price the vendor states. Where a measured deduction gives a different rate (Wan: 6 credits for 2 s, 15 for 5 s, 30 for 10 s, i.e. 3 credits per second), both figures are kept side by side and neither is averaged. Nothing is interpolated, and no model gets a cost figure its vendor did not publish. - No "best model" ranking. This is a record of terms, not a benchmark.
- No scraped marketing copy. Where a vendor's own wording matters, it is quoted, not paraphrased into something stronger.
Source and updates
Maintained alongside videofreetier.com — the same figures appear there with the surrounding context, and corrections land on that page first. The dataset is the machine-readable half; the site is the readable half.
Mirror: github.com/nicoleiszsy/seedance-2-5-free-tier-data
Contributing
If a figure is out of date or wrong, open a discussion or a pull request with the correct value and the date you observed it. A correction carrying a date is worth more than the original row. Please do not send a value without a date — it cannot be merged into a dated dataset.
Licence
CC BY 4.0. Use it, quote it, build on it; credit VideoFreeTier and link back to https://videofreetier.com/.
Structure
default 74 rows
| Split | Rows | Size |
|---|---|---|
| train | 74 | 19.9 KB |
Details
- Repository
- videofreetier/seedance-2-5-free-tier-data
- Publisher
- Zhou
- Task category
- Not stated by the source
- Tags
- ai-video, seedance, free-tier
- Size category
- n<1K
- Languages
- en
- Revision
- 509fb83923856387566cd1467b29a55489ccc6da
- Last updated
- 2026-10-04
Files
6 files, 87.5 KB in total.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| seedance-2-5-free-tiers.csv | Data | 16.1 KB | — |
| seedance-2-5-free-tiers.json | Data | 27.2 KB | — |
| LICENSE | Documentation | 18.7 KB | — |
| README.md | Documentation | 5.4 KB | — |
| seedance-2-5-free-tiers.md | Documentation | 17.7 KB | — |
| .gitattributes | Repository | 2.5 KB | — |
License and Download
- License
- cc-by-4.0
- Access
- No access gate
Released by Zhou through its official repository on Hugging Face. Read the license.