Sora 2 Free: Three Different Questions, and How to Check Each
We have not measured Sora 2's free tier, so this page quotes no Sora 2 allowance numbers. What it gives you instead is the instrument this site uses to measure a free tier, applied to the three different things people mean by "sora 2 free": a recurring free allowance, a time-boxed trial, and free access to a Pro model. Run it yourself and you get a number you can defend. Guessing gets you a number that changes next month.
Nothing here requires a sign-up. Every dated figure this site holds is published as CSV, JSON and Markdown on the [data page](/data/) — no account, no email gate, nothing to register for before you can read a number.
The three questions behind one phrase
Searchers typing this phrase are usually asking one of three different things, and the three answers do not substitute for each other:
| The phrase | What is actually being asked | What kind of answer it gets |
|---|---|---|
| sora 2 free | Is there an ongoing allowance that refills? | A quota with a reset period |
| sora 2 free trial | Is there a one-off grant that ends? | An amount plus an expiry date |
| sora 2 pro free | Is the higher-end model reachable without paying? | Which tier unlocks which model |
A recurring allowance and a trial look identical on the day you receive both. They diverge a week later: one refills, the other is gone. Any answer that reports an amount without saying which of the three it is has not answered the question.
This site holds dated observations for one model — Seedance 2.5 — on named channels. It holds none for Sora 2. Rather than borrow a figure from someone else's page and republish it as fact, the rest of this page shows how the measurement is done, so you can take it yourself and record it with a date attached.
The instrument: six fields per free tier
Every row in this site's dataset carries the same six fields. That is the whole method — there is no estimation step anywhere in it.
- Model — which model the grant applies to, named exactly.
- Channel — where the grant was obtained. The same model grants different amounts on different channels, so a number without a channel is not a fact about the model.
- Metric — what the grant is measured in, with the unit written out.
- Value — what was granted, as stated or as observed.
- Observed — the date it was seen. Free-tier terms move without announcement; the date is what makes a number checkable.
- Source type — whether the vendor stated it, a user reported it, or a single uncorroborated account claims it.
Six fields, no arithmetic. The discipline is refusing to fill in a field you do not have.
Worked example: the same six fields on a model we did measure
Here is the instrument running on Seedance 2.5, the one model this site has dated observations for. These are the actual rows, not an illustration:
| Channel | What was granted | Unit | Date observed | Source type |
|---|---|---|---|---|
| Jimeng | 60–100 per day | credits | 2026-07-04 | report |
| Jimeng | 88 per day | credits | 2026-08-11 | report |
| Doubao | 5 per account per day | generations | 2026-08-11 | report |
| Mina Labs | 20 on sign-up, one-off | credits | 2026-08-11 | report |
| Lovart | up to 20 generations, tied to an annual plan | generations | 2026-08-11 | report |
| Doubao Work | 30 days of Standard membership (list CNY 68) | time, not output | 2026-08-25 | report |
Read that table and the shape of any free tier becomes visible. Five channels, one model, and no two of them grant the same thing: some grant output, one grants *time*, one grants a one-off that never returns. Anyone who asks "how much do you get free" without naming the channel is asking a question with five correct answers.
What we do when two rows disagree
The two Jimeng rows above conflict — 60–100 per day on 2026-07-04, and 88 per day on 2026-08-11. Both are kept, both are dated, and neither replaces the other.
That is the rule, and it is worth copying, because the tempting move is to average them or pick the newer one. Averaging invents a number nobody stated. Picking the newer one asserts that the terms changed, which is a separate claim needing its own evidence. Keeping both with their dates and marking the conflict unresolved says only what was observed.
The same rule covers the API availability rows: the vendor says "coming soon" (ByteDance official, 2026-07-31) while a developer report says it opened on 2026-07-16 (reported 2026-07-15). Both are in the dataset, the conflict is flagged, and the dataset does not adjudicate it.
Four checks that catch most bad numbers
These are the checks that have mattered most on the model we did measure. Each one is a question to ask of whatever figure you are looking at.
Check the unit before the amount. Credits, generations and membership duration are three different currencies with no published exchange rate between them. A grant of "20" is not comparable to a grant of "20" until you know what the 20 counts — and a grant measured in days is not comparable to either, because it meters your calendar rather than your output.
Ask whether the grant refills. A one-off sign-up bonus and a daily allowance are both free, and only one of them is still there tomorrow. Mina Labs granted 20 credits once on 2026-08-11; Doubao granted 5 generations per account per day on the same date. Same day, same model, opposite behaviour over time.
Find out what happens to output you already made. This is the check people skip, and it is the one that changes plans. On the channel observed, free-tier output is watermarked, removal is off by default even on the paid tier, and it is not retroactive — subscribing later does not strip watermarks from clips already saved (Doubao, 2026-08-27). Whatever a free tier grants you, the question that decides its value is whether what it produced can ever leave it.
Separate the grant from the yield. The amount you are given is not the amount you keep. The only first-attempt usable rate measured in this dataset is roughly 30% to over 60% — one host, five test groups, one tester, on 2026-08-12, with prompts that assigned one job per reference. No vendor publishes that conversion rate, and it is the single number that turns an allowance into a schedule. Measure your own; do not borrow this one.
Applying it to Sora 2 in ten minutes
If you want a defensible answer for Sora 2, this is the shortest path:
- Open the official Sora page and record what it states about free access, with the date you read it.
- Record the unit — generations, credits, or time — before the amount.
- Determine whether it resets daily, monthly, or never.
- Check whether the figure you found is a trial with an end date rather than a standing allowance.
- Check separately whether "Pro" names a different model or a paid tier, and which one the free access reaches.
- Generate one clip and record whether it carries a watermark, and whether paying later removes it from clips already saved.
- Write all of it down with today's date, and re-check in a month.
Step 6 is the one that decides whether the answer matters. Everything else is accounting.
Scope of what this page says
The dated figures above are observations of Seedance 2.5 on Jimeng, Doubao, Doubao Work, Mina Labs, Lovart and BytePlus ModelArk between July and August 2026. They are not statements about Sora 2, they are not industry-wide findings, and they do not predict what any channel offers next month — terms moved between July and August 2026 without announcement on the one model we watched.
The complete dataset — 22 rows, each carrying its channel, observation date and source type — is on the [data page](/data/) in three formats, free to use under CC BY 4.0 with attribution. To turn a grant and a usable rate into an attempt count for your own piece, the [attempts-needed calculator](/tools/free-tier-calculator) uses only figures from that dataset and estimates nothing on your behalf.