2026-10-05 · VideoFreeTier

Luma Dream Machine Free: About 30 Generations a Month, Never Printed on the Price List

Luma Dream Machine's free tier is a monthly allowance of about 30 generations — reset each month, not carried over — and one generation costs one unit of it. A free account runs at Draft resolution only: 540p, 720p and above are not available. Every output carries a watermark that cannot be removed, and the licence is personal and non-commercial. Video extension and local repaint are not available either. That is the whole free tier, and none of it is printed on Luma's pricing page.

The plan table on lumalabs.ai/pricing lists three individual plans — Plus at $30/month with 10,000 credits, Pro at $90/month with 40,000, Ultra at $300/month with 150,000 — and two business plans, Team and Enterprise. There is no free row. The word "free" does appear on that page, eleven times in its source, in four distinct strings, and the only one of them that says a free grant exists is "All plans include free trial credits and commercial use" — a sentence in the page's structured data, written for machines rather than for readers, with no amount attached. Luma's own FAQ, the item titled "How does pricing work?", names the same three plans and no fourth.

So the number exists and the page does not have it. The grade matters as much as the number here. The nine vendor rows this site holds on Luma say nothing about a free allowance. Every row that describes one is a report row — five reported by the site owner from a logged-in account on 2026-10-05, and one measured here. The allowance is reported as approximate, no screenshot is on file for it, and this site has not independently re-verified it.

Watch the same finding as a video: youtu.be/9CctZhdmxJU

Nothing here requires a sign-up. Every dated figure this site holds is published as CSV, JSON and Markdown on the data page — no account, no email gate, nothing to register for before you can read a number.

What the free tier actually is

Six rows describe it. All six are dated 2026-10-05, and all six are reports rather than vendor statements:

Read the third column before the second. Every figure in that table was reported by the site owner from a logged-in account, with no screenshot on file and no independent re-verification — which is a weaker grade than a number read off a vendor's own page, and a stronger one than a guide repeating what a guide said. The access row is the only one this site measured directly: an unauthenticated request to app.lumalabs.ai was redirected to an authentication sign-up page and returned HTTP 403 behind a bot check.

Two details in that table are worth pulling out, because they are the ones a reader planning a project will trip over.

The allowance is counted in generations, not credits. One generation costs one unit, whatever the model — so the number that matters is not "how many credits do I have" but "how many attempts do I get". That is the same shape as Grok Imagine's free tier in this dataset and a different shape from every credit-balance model in it, where the same grant converts into wildly different amounts of video depending on the model you pick. On Luma's free tier the model choice does not change the price of an attempt. It changes only what the attempt can look like.

A monthly allowance is a rate, not a stock. It refills, so a project can be planned across months — and because it does not carry over, an unused month is spent rather than banked. The no-rollover behaviour is the one part of that table a vendor page corroborates, from the other direction: Luma's own FAQ says of paid plans that "any unused credits expire at the end of the applicable billing cycle and will not be carried over to the subsequent month." The free tier is reported to behave the same way, and that report is consistent with the vendor's stated rule rather than in tension with it.

The free tier's ceiling is the cheapest rung of the price ladder

Here is where the pricing page and the free tier finally touch.

Luma's pricing page prices its own model, Ray, in detail — and it prices it twice, once for Ray 3.2 and once for Ray 3.14:

Ray 3.14's column is stated by the page as per-second rates. Ray 3.2's per-second column is our division of the page's own five-second figures — 20, 50, 100 and 400 all divide by five to exactly 4, 10, 20 and 80. The two columns are the same four numbers, so across two model versions the base ladder does not move at all; what the page's own header changes between them is the HDR multiplier, 2× for Ray 3.2 and 4× for Ray 3.14.

Now put the reported free tier beside it. The free account's ceiling is Draft, and Draft is the bottom rung of that ladder — the cheapest thing Luma sells, at 20 credits per five seconds, or four credits a second. The most the free tier can produce is priced at the least the paid tiers can buy. The page's ladder runs upward from exactly the point the free tier stops, and 540p — the first rung above it — is the first thing a free account cannot reach.

Applying the multiplier the header states, the top of the Ray 3.2 ladder is not 400 credits per five seconds but 1,200 for 1080p HDR+EXR, sixty times the Draft rate. That is our arithmetic on the page's stated multiplier rather than a price the page prints, and it is the kind of figure that decides whether a free grant is worth claiming: at about 30 generations a month, the free tier is not competing with that end of the ladder at all.

Eleven occurrences of the word "free"

On 2026-10-05 the source of lumalabs.ai/pricing contained eleven occurrences of the word "free". They resolve to four distinct strings, and this table is all of them:

Read the third and fourth rows first. Six of the eleven occurrences sit inside names no visitor will ever see — an experiment flag and a campaign field in the page's own configuration. They are not disclosure; they are plumbing, and they are counted here only because the count is the point.

That leaves two sentences. One is visible: the MiniMax H3 Max row advertises that a generation's first two to four reference images cost nothing, then about seven to thirteen credits each after that. It is the only place on the page where a reader meets the word "free", and what it describes is free input to a generation that is itself priced — on a third-party model, MiniMax H3 Max, that Luma hosts rather than builds. The other sentence is not visible at all. It sits in the JSON-LD block the page publishes so that search engines and agents can describe it.

So the pricing page tells a machine that free credits exist, and tells a reader nothing about them. That is the whole of Luma's public free-tier disclosure on the page a search for "luma dream machine free" is most likely to land on: a machine-readable claim with no number in it — and, as the rows above show, a real allowance of about 30 generations a month that a reader has to log in to find.

The page that answers "How does pricing work?"

Luma's Learning Center publishes a FAQ item titled "How does pricing work?". Its published answer reads:

Luma runs on a monthly subscription with a credit allowance included in each plan. Credits are used every time you generate images, video, audio, with the amount varying based on what you're creating. There are three plans to choose from: Luma Plus at $30/month includes 10,000 credits; Luma Pro at $90/month includes 40,000 credits; Luma Ultra at $300/month includes 150,000 credits.

Read the sentence that is not about money. A FAQ whose whole job is to explain how pricing works describes the product as a monthly subscription with an allowance in each plan, enumerates three plans, and stops. On a site about free tiers that would be a finding. On a page whose question is literally how pricing works, it is the answer.

Two more rows come from the same FAQ, and one of them turned out to be the closest thing to a published description of how the free tier behaves:

• Unused credits expire. The FAQ's own wording: "Please note that any unused credits expire at the end of the applicable billing cycle and will not be carried over to the subsequent month."
• Failed generations are refunded. "Any credits used for failed generations will be automatically returned to your account."

The second is the opposite of a pattern this dataset already holds. On Seedance 2.5's API, platform failures were not charged but "dissatisfaction" was — and a system cannot detect "it does not match what I asked for", so that case always bills. Luma states a flat refund rule with no qualification. It is recorded here as the vendor's wording, not as a behaviour this site has tested.

What this site could measure, and what it could not

The Luma block in this dataset now runs to fifteen rows, dated 2026-10-05. Nine are vendor statements read from Luma's own pages. Six are reports, and they are not all the same kind of report:

Measured here. One row is an observation this site took directly. An unauthenticated request to app.lumalabs.ai was redirected to auth.lumalabs.ai/sign-up carrying a client_id and a redirect_uri, and the response was HTTP 403 behind a bot-check interstitial. No generation path was reachable without an account. That is the same shape as Wan 2.5's login gate and Kling AI's redirect: recorded because the absence of a path is a fact, and because it is the step that stood between this dataset and the numbers above.

Reported, not measured. Five rows are what the site owner read from a logged-in free account: the allowance, the generation-count mechanism, the Draft-only resolution ceiling, the watermark and licence, and the missing editing features. Those rows are graded report, they carry no screenshot, and the allowance is stated as approximate rather than exact. They are the best evidence this dataset has on Luma's free tier and they are not the same grade as a figure printed on a vendor's page — which is why the article leads with the grade and not only with the number.

Still missing. No Luma generation has been run and recorded before and after by this site, so there is no measured deduction: nothing that shows the balance moving by one unit, and nothing that would test the reported one-generation-one-unit rule against a real account. That row does not exist, and this page will not fill the gap by treating the reported allowance as if it had been measured. When it does exist it will look like the others in this dataset — its own row, its own date, a note beginning site-measured: — exactly as PixVerse's 50-credit deduction and Wan's 6 / 15 / 30-credit deductions for two, five and ten seconds carry theirs.

Luma also joins a short list here of models whose free-tier record opened with absences rather than numbers. PixVerse was the first, with an official marketing site carrying no free-tier credit figure at all. Wan 2.5's free column was the second, sitting beside a fully priced cost table on the same page. Luma was the third, and it is the first where the only official sentence confirming that a free grant exists is one a human reader never sees — and the first where a number arrived from a logged-in account before any page published one.

The record now: 99 rows across eleven models

Luma's five new rows land the dataset at 99 rows across eleven models — Seedance 2.5 (23), Wan 2.5 (18), Luma Dream Machine (16), Runway (10), PixVerse (9), Kling AI (7), Veo 3.1 (7), Grok Imagine (5), Sora 2 (3), Hailuo AI (3), Canva (1). Forty-eight of those are vendor statements, forty-four are reports, and one is a single uncorroborated account. Thirteen of the reports are observations this site took itself; the rest, including Luma's five, were reported to it. Thirteen rows are a documented absence, and two of those thirteen are Luma's.

Scope of what this page says

The nine vendor Luma rows were read from lumalabs.ai/pricing on 2026-10-05 — the plan cards, the page's structured data and its per-model cost table — and from lumalabs.ai/learning-center on the same date, where the "How does pricing work?" FAQ item and its published answer sit. The login-wall row was measured on 2026-10-05 in an unauthenticated session against app.lumalabs.ai. The five free-tier rows were reported by the site owner from a logged-in free account on 2026-10-05, with no screenshot on file. They describe those screens on those dates, not Luma's product in general, and the file names on this site still carry the seedance-2-5-free-tiers stem for link stability — the dataset itself covers every model named above.

Comparisons on this page are made against dated rows from the same dataset: Runway's 125-credit free deposit and its per-model cost table (runwayml.com, 2026-10-02 and 2026-10-04), PixVerse's 60/30 figures and its measured 50-credit deduction (app.pixverse.ai, 2026-10-02), Wan 2.5's plan-table rows and its measured 6 / 15 / 30-credit deductions (create.wan.video, 2026-10-03), Kling AI's 66-credit trial package with its 2026-11-03 expiry (klingai.com, 2026-10-03), Grok Imagine's count-based free tier (grok.com, 2026-10-04) and Seedance 2.5's API billing row (2026-08-04). The counts on this page — 99 rows across eleven models, fifteen Luma rows, thirteen measured rows — are counts of this site's own records, not claims about how free tiers behave in general.

Not confirmed on this page: the exact size of the free allowance, which is reported as about 30 generations a month and carries no screenshot; whether the allowance is capped per day inside the month, which is not reported; whether the Draft ceiling is a hard block or a degraded output on other settings, which is reported as unavailable rather than tested; whether the one-generation-one-unit rule holds for every model on the free tier or only the default one; whether the watermark and non-commercial terms are stated in writing anywhere on a Luma page this site has read, since the rows describing them are reports; what the free tier does with a failed generation, which the paid FAQ answers and the free tier does not; and the exact count of eleven occurrences of the word "free" on the pricing page, which is a reading of one page's source on one date and will move whenever that page is rebuilt. The free trial credits statement itself is read from structured data published by Luma and has not been independently re-verified by a second party.

The full dataset — 99 rows, each carrying channel, observation date and source type — is on the data page 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, the attempts-needed calculator uses only figures from that dataset and estimates nothing on your behalf. Related pages: PixVerse AI Free for the first model whose free-tier record was an absence, and Wan 2.5 Free for a plan page that prices every video and none of the free credits.

Update cadence: figures on this page change only when the dataset gains a dated row that touches them; every figure keeps its own observation date — check it before quoting as current.