2026-09-28 · VideoFreeTier

Grok Imagine Free: How Long a Free-Tier Number Stays True

We have not measured Grok Imagine's free tier, so this page quotes no allowance figures for it. What this site can tell you instead is something more useful than a borrowed number: on the one model we did watch, the same channel's free-tier figure changed inside 38 days, with no announcement. Any free-tier number you read — for Grok Imagine or anything else — is a reading with a date attached, and the date is the part that decides whether it is still true.

Nothing on this site needs a sign-up. Every dated figure we hold 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 38 days between two readings of the same thing

This is the observation that shapes everything else on this page. It comes from this site's own dataset, and both rows are for the same model, the same channel, and the same metric:

Same model. Same channel. Same metric. Thirty-eight days apart. The dataset records both and adjudicates neither — the note on the second row says only that it conflicts with the first and that no official statement accompanies the change.

That is the empirical basis for treating a free-tier figure as perishable. This is not a suspicion; it is a measured interval on a real product, and the interval is about five weeks.

A range became a single value — and we cannot say why

There is a detail in those two rows worth looking at directly. The first is a range, 60–100. The second is a point, 88. And 88 sits inside the earlier range.

Two readings are consistent with that, and the dataset does not choose between them:

• The underlying allowance never changed, and what changed was the precision of the report — an uncertain range narrows to a number once someone counts properly.
• The allowance did change, settling somewhere in the band it previously wandered across.

Both explanations fit the two rows equally well. Distinguishing them would need a third observation, which this site does not have. What can be said without guessing is narrower but defensible: a figure reported as a range and a figure reported as a point are not the same kind of claim, and a page that prints the second as though it replaced the first has quietly upgraded a report into a fact.

The change nobody announced

The 38-day gap is not the only interval in this dataset. Here is a second one, on a different question:

Sixteen days separate a vendor statement from a developer's account, and they contradict each other. Both rows are kept, the conflict is flagged, and the dataset takes no side.

What matters for anyone checking a free tier is the pattern rather than the verdict: in both cases, the state of the product moved across a gap of weeks, and in neither case did a public announcement mark the move. A free tier can change while every page describing it stays exactly as it was written. The pages are not lying — they are simply dated earlier than the change.

Failure modes moved too, in 21 days

Overruns are not the only thing that drifts. On the API side, what the system does when it refuses you also changed within the observation window:

• 2026-08-04 — error codes seen: 429 (rate cap), 400 (parameter limits), 402 (insufficient balance). Billing on failure: moderation and platform failures not charged, dissatisfaction charged.
• 2026-08-25 — error strings seen: api error: 529 overloaded, and connections lost mid-response.

Twenty-one days apart, and the second set is not a relabelling of the first. A 429 and a 529 overloaded ask for different responses from you: one says slow down, the other says the service is struggling. Anyone still debugging against the August 4 list on August 25 would be reading the wrong map.

The same applies to output characteristics. The 30-second single-pass clip ceiling was stated by the vendor on 2026-07-31, while the watermark behaviour — free-tier output watermarked, paid-tier removal off by default, and not retroactive for clips already saved — was reported on Doubao on 2026-08-27. Different dates, different questions, and neither is safely inferred from the other.

Three rules for dating a number you did not measure

Applied to Grok Imagine, this is what the intervals above imply. None of these rules requires a figure for Grok Imagine, and none of them invents one.

1. A number without a date is not a claim, it is a rumour. The two Jimeng rows are only usable because each carries 2026-07-04 or 2026-08-11. Strip the dates and you get "60–100 or 88", which is nonsense. Whatever allowance figure you find for Grok Imagine, the first thing to record is not the amount — it is the day you read it.

2. Staleness is not error. The 2026-07-04 reading was not wrong when it was taken. It is unverified now. That distinction matters: a figure does not become false with age, it becomes uncheckable, and uncheckable figures should be re-measured rather than discarded or averaged with newer ones.

3. Re-check on the interval you observed, not on a schedule you invented. Thirty-eight days is the longest unannounced drift this dataset captured. Re-reading a free-tier page quarterly assumes a stability that this data does not show. Monthly is the conservative cadence; anything longer is a guess about someone else's release cycle.

How to stamp your own date onto it

If you want a Grok Imagine answer you can defend next month, this is the shortest honest path — and note that most of the work is recording, not searching:

1. Find what the official source states about free access for Grok Imagine, and write down today's date next to it.
2. Record the unit before the amount — credits, generations, or a duration. "20" is not a number until you know what it counts.
3. Note whether the grant refills or is one-off. For contrast: on 2026-08-11 the same model granted 5 generations per account per day on Doubao and 20 credits once on Mina Labs. Same day, opposite behaviour over time.
4. Note whether the grant is tied to something else — Lovart's "up to 20 generations" on 2026-08-11 was tied to an annual plan, and Doubao Work's 30-day Standard membership on 2026-08-25 metered the calendar rather than the output.
5. Generate one clip and record whether it carries a watermark, and whether paying later strips it from clips you already saved. On the channel observed (Doubao, 2026-08-27) it does not.
6. Write the whole thing as one row with a date field, and re-read the official source in 30 days. If the number moved, add a new row — do not edit the old one.

Step 6 is what separates a measurement from a screenshot. [Sora 2 Free](/posts/sora-2-free/) covers the six fields to fill in for a first measurement of any model; this page covers what to do after you have one.

Scope of what this page says

Every dated figure above is an observation of Seedance 2.5 on Jimeng, Doubao, Doubao Work, Mina Labs, Lovart and BytePlus ModelArk between July and August 2026. None of them describes Grok Imagine, none is an industry-wide finding, and the 38-day interval is a single measured case, not an average or a forecast — one model, one channel, one metric, twice.

The full dataset — 22 rows, each carrying 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 an allowance and a usable rate into an attempt count, the [attempts-needed calculator](/tools/free-tier-calculator) uses only figures from that dataset and estimates nothing on your behalf.