Luma

Where Luma sits

First, a naming note, because half the reviews of this product are now describing something that no longer exists: Dream Machine is retired as a product name. The old lumalabs.ai/dream-machine address redirects to the app, and Luma now presents itself as a creative workspace with agents — one subscription that runs its own Ray3.2 video and Uni-1 image models alongside a catalogue of third-party models including Veo 3.1, Kling 3.0, Seedance 2.0 and 2.5, Minimax H3, Flux 3, GPT Image 2, Nano Banana 2 and Pro, Seedream, and ElevenLabs audio models. The pitch is that an agent takes your brief and picks the right model for each step, instead of you shopping model by model. Whether that bundling is worth it depends entirely on how many of those models you would otherwise pay for separately.

What I look for in this category is output quality per dollar and whether the pricing punishes iteration, because nobody in video generation gets a usable take on the first try. Luma scores well on the first and has complicated feelings about the second.

What it actually does

The core is text-to-video and image-to-video on Ray3.2, generating 5 or 10 second clips at up to 1080p, with 16-bit HDR output that one testing outfit describes as the only model in its class to do so — though Ray3.2 has no native audio, so sound is bolted on from the ElevenLabs models or the Kling/Seedance/Veo routes that do carry audio. Ray3.14, the January 2026 iteration, is the budget workhorse: 4x faster and 3x cheaper at 720p than base Ray 3, at 20 credits per second at 720p versus Ray3.2's heavier pricing. Luma describes Ray 3 as a reasoning model that evaluates and retries its own output before showing you a result, which is a direct answer to the credit-burn complaint. Beyond generation there is Modify (video-to-video editing from 180 credits for 5 seconds at 360p), Reframe (aspect-ratio changes at 10 to 120 credits per second), keyframe control, and the agent layer that composes multi-step projects.

The credit table is where this product lives or dies. At time of writing, from the pricing page: Plus is $30 a month billed monthly or $25 a month billed yearly ($300 up front) for 10,000 credits; Pro is $90 or $75 ($900) for 40,000 credits and 4x agent usage; Ultra is $300 or $250 ($3,000) for 150,000 credits and 15x agent usage; Team and Enterprise are quote-only with shared credits, SSO, and analytics. There is no free tier on the pricing page — the page mentions free trial credits without a number. Credits renew monthly and do not roll over, though Luma refunds credits for generations that error out, and $4 top-up packs of 1,200 credits exist for subscribers and roll over for 12 months.

The unit costs vary by more than an order of magnitude across models, which makes the plan allowance almost meaningless in the abstract. Ray3.2 in standard dynamic range costs 20 credits per 5 seconds at Draft quality, 50 at 540p, 100 at 720p, and 400 at 1080p; HDR doubles that and HDR plus EXR triples it. A 10-second clip costs exactly three times a 5-second one (1,200 credits at 1080p). Third-party models run on the same credits: Kling 3.0 at 1080p with audio is 49 credits per second, Seedance 2.5 at 720p is 78, Veo 3.1 with audio is 280. Translate that to dollars on Plus ($3 per 1,000 credits) and one 5-second 1080p Ray3.2 take costs $1.20, while a 10-second Veo 3.1 clip costs $8.40. A separate API exists with per-clip pricing that works out similar at full pool utilization — $0.30 per 5-second 720p Ray3.2 clip, $1.20 at 1080p — with a friendly refund policy on failed generations.

My testing notes

Three observations. First, budget discipline matters more here than in any rival I test. 1080p costs eight times 540p on Ray3.2, and HDR doubles again. The workflow that keeps a Plus subscription sane is drafting at 540p or Draft quality — where 10,000 credits buys 200 to 500 clips — and only promoting approved shots to 1080p. One comparison site calculated that a finished 30-second 1080p ad runs about $7.20 per take at list price and closer to $21.60 once you assume the three takes per shot that reality demands. That is not a scandal; it is just the economics you should model before paying.

Second, the product around the product is in churn. The Dream Machine rebrand is recent enough that iOS in-app purchases still carry old plan names (Lite, Unlimited) that no longer match the web tiers (Plus, Pro, Ultra), so buying through the App Store can land you on a plan the website no longer describes — check exactly what the app is selling before you tap. Worse for Android users: there is no official Luma video app on Google Play at all. The app called 'Luma Dream Machine' there is published by ApexAI Tech Limited, is not Luma, and charges $9.99 a week. I would call that out as an actively harmful marketplace confusion for anyone searching for this product on Android. On the web, though, the workspace itself held up: in a third-party head-to-head test on a Warsaw tram scene, Ray3.2 produced the most atmospheric shot of the models tried, and the 16-bit HDR path is a real differentiator for grading.

Third, the credit refund policy is genuinely consumer-friendly and I want to name it: generations that end in error get their credits back, top-up credits survive cancellation and roll over for 12 months, and commercial-use rights cover material made during your subscription even after you cancel. Compare that to the subscription-credit status quo and Luma is treating iteration risk seriously on the edges, even if the monthly credits themselves expire.

What it does not do

There is no free tier and no unlimited option — the old Dream Machine Unlimited plan has disappeared from the site, and any comparison quoting Lite or Unlimited is reading pre-2026 data. Ray3.2 clips have no native audio, so sound is a separate step through other models. Clips cap at 10 seconds per generation, so anything longer is an editing exercise. Monthly credits do not roll over. There is no official Android video app. The agent layer's real value is hard to separate from the bundle pricing — if you only ever use one model, you are paying Luma's margin for a catalogue you do not touch. And as with every model in this class, hands and physics still fail on a meaningful minority of takes; one reviewer's estimate of needing five generations per usable clip is on the pessimistic side but not fantasy.

Verdict

Luma is the most credible single subscription for mixed-model creative video work right now, and the combination of Ray3.2's HDR output, cheap Ray3.14 drafting, and third-party models on one credit balance is genuinely useful if you would otherwise hold two or three separate subscriptions. I rate it 4.1 for depth of core functionality across video, image, and audio, mixed on value — good per-clip economics at the draft tier, punishing at 1080p HDR with Veo — strong web usability undercut by the Android absence and App Store plan-name drift, a real API and help ecosystem, and reliability signals split between a fair refund policy and obvious product-line churn. Draft at 540p, buy top-up credits when you run close to the line, and verify what the iOS app is selling you before you subscribe.

Sources

• https://lumalabs.ai/pricing
• https://quantslant.com/luma-ai-dream-machine-ray32-review-pricing-test
• https://quantslant.com/luma-ai-pricing
• https://digijaws.com/reviews/luma-dream-machine-review
• https://eesel.ai/blog/luma-ai-pricing
• https://hooked.so/compare/luma-ai-pricing
• https://adtestbench.com/tools/luma
• https://ugcgen.ai/luma-ai-pricing
• https://docs.lumalabs.ai/
• https://yesilcamstudios.com/compare/yesilcam-vs-luma