This review is built from what could be checked on November 3, 2026: Luma's pricing, Ray, UNI-1, app and API pages, its API documentation, news posts and learning center; its terms, privacy, billing, moderation and API policies and its EU training-data summaries; Artificial Analysis's video and image leaderboards; a competitor's evaluation published by Black Forest Labs; reporting by Advanced Television, The AI Economy, CairoScene, Post Magazine, PPC Land, IT Brief and MoguraVR; the App Store listing and Apple's feed of recent US reviews; and rivals' own pricing pages. Luma's claims about itself are labelled as its own.
What Luma AI is
Luma AI is the brand of Luma AI, Inc., a company founded in 2021 by Amit Jain and Alex Yu and based in Redwood City, California (about page). It began with 3D capture, launched the Dream Machine video app in June 2024, and now sells three things: the Luma app, a browser workspace where creative agents plan and generate across image, video, audio and text; its own models, Ray for video and Uni for images; and an API for developers.
Most buyers meet it as the Luma app at app.lumalabs.ai, with an iOS app still listed as Luma Dream Machine. One credit balance pays for Luma's own models and for other labs' models:
| Who made it | Models on Luma's pricing page (November 3, 2026) |
|---|
| Luma | Ray3.2 and Ray3.14 (video), Uni-1 (images), upscaling, HDR conversion |
| ByteDance | Seedance 2.5, Seedance 2.0, Seedance 2.0 Mini, Seedream 5.0, Seed Audio |
| Kuaishou | Kling 4.0, Kling 4.0 Omni, Kling 3.0, Kling 3.0 Omni and Turbo |
| Google | Veo 3.1, Gemini Omni Flash, Nano Banana Pro, Nano Banana 2 and 2.1 |
| OpenAI | GPT Image 2, GPT Image 2.5 Flare and Sunburst, GPT Image 1.5 |
| Others | MiniMax H3, Flux 3, Ideogram 4.5, ElevenLabs v3 and Music, Sync lip-sync, Topaz upscaling |
The agent picks a model for each step unless you choose one, and Luma's own FAQ warns that it can misread instructions.
What changed in 2026
The biggest change is that Dream Machine is no longer the product. On March 5, 2026 Luma launched Luma Agents and its first Uni image model; Post Magazine and PPC Land reported the launch with Publicis Groupe Middle East and Serviceplan as early agency users (Post Magazine). New users now start in the agent app, Dream Machine stays open only to existing users, and its boards do not migrate automatically (Luma). Photon, the image model behind Dream Machine, has been retired.
Since then Luma has shipped a steady run of features:
- Models: Ray3.14 with native 1080p (January 26), the Uni-1.1 API (May 5), and Ray3.2 with the first full Ray API on June 9, which IT Brief and MoguraVR covered (IT Brief).
- Hosted models: Seedance 2.5 and MiniMax H3 (August 6) and Kling 4.0 (October 23).
- Workflow tools: team members and admins (April 1), Skills (June 16), Layers (July 29), Scenes (August 11), Ad Variants (October 1) and eight one-click mini-apps (October 27).
- Elsewhere: Claude Motion animations open in Luma (October 8), and Luma plugins for ChatGPT and Claude spend your Luma credits from the chat (guides published October 19).
As of November 3, Luma's app also announces Character Sheet, which keeps a character consistent across views, and Expression Sheet, which shows "your character's emotional range in one reference sheet" (Luma's site). Luma had published no dated post about either.
The company behind it has grown fast. Luma raised a US$900 million Series C on November 19, 2025, led by the Saudi AI company HUMAIN, with AMD Ventures, Andreessen Horowitz, Amplify Partners and Matrix Partners (Advanced Television). The AI Economy reported that the valuation was not disclosed; CairoScene put it at about US$4 billion. Luma says it has raised more than US$1.1 billion in all, opened offices in London and Riyadh, and in June 2026 started an Open Physical AI Lab for robotics, with access to HUMAIN's planned 2-gigawatt Project Halo cluster. It also now sells hosted GLM 5.2 text inference to LLM routers and resellers, a separate product with its own keys.
Plans and what credits buy
| Plan | Price | Credits a month | Cost a credit | Ray3.2, 5 s at 720p | Notes |
|---|
| Plus | US$30 a month, or US$300 a year | 10,000 | 0.3 cents | 100 clips | Commercial use; guest editors |
| Pro | US$90 a month, or US$900 a year | 40,000 | 0.225 cents | 400 clips | 4x usage with Luma Agents |
| Ultra | US$300 a month, or US$3,000 a year | 150,000 | 0.2 cents | 1,500 clips | 15x usage with Luma Agents |
| Team | Contact sales | Shared | Custom | Custom | Team management, usage analytics, SSO |
| Enterprise | Contact sales | Custom | Custom | Custom | Custom fine-tuning; no-train guarantee in its terms |
Source: Luma's pricing page. There is no free plan on it. Plus's 10,000 credits buy 100 five-second Ray3.2 clips at 720p, 25 at 1080p, 500 drafts at 360p or 333 Uni-1 images at 30 credits each.
Four rules decide what that money is worth (billing policy):
- Credits expire. Unused plan credits end with the billing cycle and do not roll over.
- Failures are free, misses are not. Credits for failed generations come back automatically; a generation that finishes but misses the brief is charged.
- Extra use is billed afterwards. Past your allowance you can set a spending limit; extra credits are billed at the end of the period and are non-refundable. The spend-limit policy names no price a credit.
- Refunds are narrow. An unused plan may get a full refund, but only for the most recent billing period, at Luma's discretion. Annual subscribers can switch to monthly and get the remaining months back after the used months are repriced at the monthly rate.
The terms add that a cancellation must arrive at least one day before renewal, and that Luma may change prices with 30 days' notice. Luma's own learning center warns that the agent may generate several variations or retry automatically unless told to make one version and wait (Stop Wasting Credits).
What each model costs on Luma
| Model | Credits | 5 seconds on Plus | Sound |
|---|
| Ray3.2, 720p | 100 per 5 s | US$0.30 | No |
| Ray3.2, 1080p | 400 per 5 s | US$1.20 | No |
| Ray3.2 edit (V2V), 720p | 72 a second | US$1.08 | Keeps source audio |
| MiniMax H3, 768p | 28 a second | US$0.42 | Yes |
| Kling 4.0, 1080p | 79 a second | US$1.19 | Yes |
| Veo 3.1 | 140 a second | US$2.10 | Yes |
| Seedance 2.5, 1080p | 191 a second | US$2.87 | Yes |
Source: Luma's pricing page; dollar figures use Plus's 0.3 cents a credit and are about a third lower on Ultra. For Ray3.2 the pricing page also notes "No keyframes: plus 30 credits". Images range from 1 credit for Seedream 4.0 to 255 for GPT Image 2 at high quality in 4K; Uni-1 is 30 and Nano Banana Pro at 2K is 35.
For hosted models Luma is competitive. A 5-second Kling 4.0 clip at 1080p costs 395 Luma credits, about US$1.19 on Plus and US$0.79 on Ultra; Runway lists the same clip at 118 credits, about US$1.18 at its Max rate, or 89 credits during a discount that ends November 30.
Ray3.2: what it does and where it stops
Ray3.2 is built for direction rather than one-shot prompts. Luma's Ray page lists:
- Keyframes: up to 16 inside one clip, so you set where the look shifts and where a movement lands.
- Modify Video V2: restyles existing footage, keeping body pose and the expressions of up to eight faces, for up to 20 seconds at 24 fps, 15 at 30 fps or 7 at 60 fps, at 1080p.
- Reframe: extends a shot to 9:16, 3:4, 1:1, 4:3, 16:9 or 21:9 with scene-aware fill.
- HDR and EXR: native 16-bit HDR and EXR frames in ACES2065-1, at two and three times the standard credits.
The limits are on the same page. Text-to-video and image-to-video run 5 or 10 seconds and make no sound. Modify Video struggles with heavy camera shake, very fast action, dark or blurry footage, crowded scenes and hidden faces, by Luma's own account. Through the API, HDR is limited to 720p or 1080p and 5 seconds, and extend and reframe work only in standard dynamic range.
Luma's EU training-data summary adds something its product pages leave out: Ray 3 was initialised from a third party's open-weight model and then trained further by Luma, and Ray3.14 and Ray3.2 derive from it (summary). Luma does not name the base model.
API prices
| Ray3.2 task | 540p | 720p | 1080p |
|---|
| Text or image to video, 5 s | US$0.15 | US$0.30 | US$1.20 |
| Text or image to video, 10 s | US$0.45 | US$0.90 | US$3.60 |
| Video edit, 10 s | US$1.08 | US$2.16 | US$4.32 |
| Reframe, per second | US$0.06 | US$0.12 | US$0.36 |
Source: Luma's API page, for standard dynamic range; HDR costs twice as much and HDR with EXR three times. The API counts in credits, with a default list rate of US$0.003 a credit, the same as Plus (API documentation). Uni-1 images cost 14 credits each and Uni-1 Max 34, about US$0.04 and US$0.10, which is cheaper than the 30 credits a Uni-1 image costs in the app.
Pay-as-you-go has rate limits and no latency guarantee. Provisioned Throughput sells dedicated capacity from US$3,800 a unit for one month down to US$2,100 a unit-month on a one-year term, with a minimum of 8 units; Luma's API page says 4. Customers of the older Dream Machine API must move from Ray 3, Ray 2, Ray 2 Flash and related models to ray-3.2 on the new API before the date in their deprecation notice.
What public tests and leaderboards show
Outside evidence on Luma's current models is thin, and Luma's terms are part of the reason: section 3.4 bars users from publishing benchmarks or performance information about the service, and the API terms require Luma's written consent. What exists:
| Board (Artificial Analysis) | Luma model | Rank | Elo | Leader | For comparison |
|---|
| Image to video, silent | Ray 3.2 | 33 of 34 | 818 | Utopai X, 1154 | Runway Gen-4.5 900, Veo 3.1 953, Kling 3.0 Pro 1000 |
| Text to image | UNI 1 Max | 33 of 106 | 1024 | GPT Image 2.5 Sunburst, 1223 | UNI 1: 51st, 993 |
| Image editing | UNI 1 Max | 25 of 77 | 1080 | GPT Image 2.5 Sunburst, 1199 | UNI 1: 40th, 1042 |
Sources: image to video, text to image and image editing, checked November 3, 2026. Ray 3.2 joined the image-to-video board at the end of October with 2,332 votes and is absent from the text-to-video boards. Several models Luma hosts, Seedance 2.5 among them, sit near the top of the same board.
- Black Forest Labs, a competitor, July 23, 2026. In preliminary tests of 10-second 720p clips with audio, raters preferred FLUX 3 over Ray 3.2 in 93% of comparisons (FLUX 3 post). The setup favours models that make sound, and Black Forest Labs also sells FLUX 3 through Luma.
- Luma's own claims. Luma's October 2025 evaluation report says third-party evaluators put Ray3 level with Veo 3 and ahead of Runway Gen-4, Midjourney Video and Moonvalley Marey. Its UNI-1 page says Uni-1 ranks first in human preference Elo overall, for style and editing and for reference-based generation; the independent boards above place it lower.
No published same-prompt comparison of Ray3.2 against Kling 4.0 or Seedance 2.5 was found.
What users report
The iOS app, still called Luma Dream Machine, rates 4.5 from 1,800 US ratings (App Store). Recent reviews are far harsher. In Apple's feed of the 50 most recent US reviews, written between January 16 and October 27, 2026, 33 gave one star and 7 gave five. Of the 38 one- and two-star reviews, 19 complain about credits, charges, price or a card being required before anything could be tried; one Plus subscriber says 10,000 credits were gone in 20 minutes without a finished video and a refund was refused. About ten describe crashes, sign-in loops or an app that would not load, and several say the agent ignored the prompt or changed faces. The five-star reviews praise the results and the speed.
Two practical warnings follow. The iOS app was last updated on April 17, 2026, while most new features arrive in the browser. And the App Store lists unrelated apps named Luma from other sellers, so check that the seller is Luma AI, Inc.
Privacy, training and content rules
- Training: the terms of service updated May 14, 2026 license Luma to use paid users' inputs and outputs to create, test, improve and train its models, and expand the license for free use to public display and distribution. The terms describe no opt-out for Plus, Pro or Ultra.
- API and Enterprise: the API terms of April 28, 2026 promise no training on API inputs or outputs, and the Enterprise terms carry a No Train Guarantee and IP indemnification, including for training data.
- What has happened so far: Luma's EU summaries of October 23, 2026 say Ray and Uni models were not trained on user prompts or uploads. Its own crawler gathered images, video and text from a hand-picked list of sites from October 2023 to April 2026 under the default python-requests user agent, with no crawler name of its own, and Luma is not a signatory of the EU's code of practice for AI models.
- Who sees your requests: the subprocessor list of October 4, 2026 names model providers including BytePlus, Kling AI and MiniMax entities in Singapore, OpenAI, Anthropic, Black Forest Labs, ElevenLabs and fal.ai, and marketing tools including Meta Pixel, LinkedIn Tag, Google Tag Manager, Apollo and Clay.
- Personal data: the privacy policy of April 20, 2026 covers uploads and agent chats, lists model training among its uses, and excludes children under 13.
- Ownership and watermarks: you own your outputs, but commercial use needs an active paid plan that includes it, and Luma may embed watermarks, content credentials or provenance data that you must not remove.
- Content: no sexual content, deepfakes, impersonation or political manipulation; accounts that keep triggering the NSFW filter are blocked, and Luma may withhold refunds from blocked accounts.
- Disputes: binding individual arbitration under California law unless you opt out by letter within 30 days.
Who should not use Luma AI
- Anyone who wants to try before paying: the pricing page lists no free plan and starts at US$30 a month.
- Buyers who judge a video model by blind-vote rank: Ray 3.2 is 33rd of 34 on Artificial Analysis's silent image-to-video board.
- People who need speech, music or sound in the same clip from Luma's own model: Ray3.2 is silent.
- Casual users who expect to bank unused credits or get money back for bad results: credits expire monthly and only unused plans are refunded.
- Self-serve users whose uploads must never train a model: choose the API or Enterprise, not the app plans.
- People who mostly use one hosted model, such as Kling 4.0: the maker's own plan may cost less a month.
Alternatives
| Option | What it is | Entry price, checked November 3, 2026 | Better than Luma when |
|---|
| Runway | Gen-4.5 and Aleph 2.0 plus Kling 4.0 and Seedance 2.5 | Free 125 one-time credits; Standard US$15 a month or US$12 billed yearly, 625 credits (Runway) | You want a free start, a cheaper plan or Aleph 2.0 for editing footage |
| Kling AI | Kling 4.0 from its maker | Standard US$6.99 the first month, then US$8.80, 660 credits (Kling) | You mainly want Kling 4.0 with sound |
| Higgsfield | Many labs' models plus its own studio tools | Starter US$19 a month, 270 credits (Higgsfield) | You want its time-limited unlimited offers, such as Seedance 2.5 |
| OpenArt | Image and video models with character tools | Starter US$14 a month, 4,000 credits; commercial use listed from Plus at US$34 (OpenArt) | You make mostly images and characters on a smaller budget |
| Pika | Video and image models on one plan | Free with credit packs; Starter US$10 a month, 900 credits (Pika) | Your budget is the smallest |
Luma is the stronger choice when you edit or reframe existing footage, need keyframe control or HDR and EXR files, or want an agent to turn one approved asset into many variants on one bill. It is the weaker one when you want the best-ranked text-to-video model, sound from the same model or a cheap way in. Credits are not comparable across sites, so run one prompt on two of them and compare the cost of the clip you keep.