Wan 3.0 vs Luma Ray 3.2: Sound Against Colour, and One Unrated Model

This is the only matchup in this series where one of the two models cannot be scored. Luma Ray 3.2 has no entry on the Artificial Analysis video arena — not on the text-to-video board, not on the image-to-video board, checked 18 September 2026. There is no blind-preference Elo for it, no sample count, no confidence interval. Anyone quoting a number for Ray 3.2 in a head-to-head is quoting a vendor-run preference test or an aggregator's own scale, and neither is comparable to the arena figures Wan 3.0 carries. So this comparison has to be made on specifications, control surfaces and billing rather than on a scoreboard — which, given how differently these two are built, turns out to be the more useful comparison anyway.

Before any of it, one naming trap that will cost you money if you miss it. Luma's version numbers do not ascend the way they read. Ray3.14 shipped on 26 January 2026; Ray3.2 shipped on 9 June 2026 and is the current flagship. The higher number is the older model. Luma has deprecated the Dream Machine branding entirely and states that Dream Machine and Ray2 are replaced by Ray — so any guide, pricing page or comparison built on "Dream Machine" is describing a product Luma no longer sells.

The missing leaderboard entry, and why it matters

Wan 3.0's standing is well documented. On Artificial Analysis's text-to-video arena with audio, captured 18 September 2026, it sits at #2 with Elo 1,229 and 6,011 recorded comparisons, behind Gemini Omni Flash at 1,233 and ahead of three other models inside a 23-point band. On the video-editing board it is #1 at Elo 1,190. Those are independent, blind, and inspectable.

Ray 3.2's standing is not documented anywhere comparable. What Luma has published is an internal "Ray3 Evaluation Report" claiming parity with Veo 3 and a lead over Runway Gen-4, Midjourney Video and Moonvalley Marey — with no numeric scores attached. Luma also claims its API runs at less than half the price and latency of comparable models, unquantified. Separately, Black Forest Labs has stated that its Flux 3 was preferred over Ray 3.2 in 93% of comparisons; that is a vendor-run test of a competitor's product and carries the same weight as Luma's own claims, which is to say not much.

This is a real gap in the record rather than a mark against the model. A model can be excellent and unrated — the arena only scores what its voters have been shown. But it changes what a comparison page should tell you. For Wan 3.0 you can say what independent voters thought. For Ray 3.2 you cannot say anything of the kind, and framing Luma's internal report or a rival's 93% claim as if either were evidence would be the single easiest way to get this article wrong.

Two different bets: sound versus colour

Here is the sharpest difference between them, and it is not about quality.

Wan 3.0 generates audio natively in the same pass as the video — dialogue, background music and sound effects, on by default, disableable with a flag, steerable with up to five reference audio clips. Alibaba's own launch materials name audio quality as an area still needing work, so treat it as a real but imperfect capability. The point is that a Wan 3.0 generation arrives as a finished audiovisual piece.

Ray 3.2 does not generate audio at all. Text-to-video and image-to-video output is silent. Audio is only preserved when you use the Modify or Reframe paths on existing footage — there is no path from a prompt to a soundtrack. This is the single largest capability difference between the two models and it is disqualifying for some workflows on its own: if your deliverable needs synchronised dialogue or sound design, Ray 3.2 is a video-only step in a pipeline that needs an audio stage bolted on afterwards, and Wan 3.0 is not.

What Luma built instead is the deepest colour and finishing pipeline of any commercial video model. Ray 3.2 generates native 16-bit HDR and exports EXR in ACES2065-1, the colour space professional grading and VFX pipelines expect. It tracks expressive facial performance across up to eight faces simultaneously, frame by frame, plus skeletal posture and gesture. It takes up to 16 keyframes per clip for frame-level direction of pacing and motion. Its Reframe feature changes aspect ratio, extends the frame or replaces a background after generation while preserving the original lighting. And it ships inside Adobe Firefly, with named studio and agency customers behind it.

That is a post-production tool. Wan 3.0 is a production tool. If you are handing shots to a colourist, the EXR export is worth more than thirty seconds of anything. If you are publishing a finished clip, silent output is a problem nothing else on the spec sheet fixes.

Duration and resolution, read carefully

Both vendors blur these numbers, in opposite directions, and both blurrings matter.

• Duration — Wan 3.0: 2 to 30 seconds in a single generation, plus forward, backward and bidirectional extension. Ray 3.2: 5 or 10 seconds for fresh text-to-video and image-to-video. Luma's widely quoted "up to 20 seconds" applies to Modify Video, its source-preserving transformation mode, not to generation from a prompt. If you are choosing on the 20-second figure, check which mode you are buying.

• Resolution — Wan 3.0: 480p, 720p and 1080p, all native. Ray 3.2: 360p, 540p, 720p and 1080p. Ray 3.2 does not do 4K. The claim that it does circulates widely and is wrong: Luma's 4K is a separate upscale utility applied after generation, not a generation mode. The HDR is real; the 4K is not.

• HDR constraints — Luma's HDR mode is limited to 5-second clips at 720p or 1080p, and HDR is incompatible with 10-second output. So the model's best colour pipeline is not available on its longest clips.

• Frame rate — Wan 3.0: 30fps. Ray 3.2: 24fps, which is the right call for film-style work and the wrong one if your delivery is 30 or 60.

• Aspect ratio — Ray 3.2 has no fixed list because Reframe adapts the output after the fact. Wan 3.0 supports 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 and adaptive at generation time.

• Inputs — Wan 3.0 accepts up to 10 reference images, 5 reference videos and 5 reference audio clips within a 20-asset ceiling, plus a document or a public web page. Ray 3.2 accepts text, images, up to 16 keyframes, and character references — a capability its predecessor Ray3.14 lacked. No documents, no URL grounding, no audio input.

What each second costs, and the billing shape underneath

Wan 3.0 is billed per second of output at Alibaba's published rates: $0.05 at 480p, $0.10 at 720p, $0.20 at 1080p. A 30-second 1080p clip is $6.00, with a limited-time 30% discount on Alibaba's own platforms running through 23 September taking that to roughly $4.20.

Ray 3.2 is billed per generation on the Luma Agents API, which makes the arithmetic different: 5 seconds at 360p is $0.06, at 540p $0.15, at 720p $0.30, at 1080p $1.20. Ten seconds doubles those figures at 360p and 540p and triples them above. At 1080p that works out to about $0.24 a second — more than Wan 3.0's 1080p rate and roughly five times its 480p rate.

Two structural notes that matter more than the per-second numbers. First, Luma's subscription credits and its API billing are separate systems and credits do not transfer between them; the consumer plans run $30, $90 and $300 a month with credits metered inside each. Budget for whichever surface you are actually building on, because the money does not move. Second, HDR and HDR+EXR multiply the credit cost — double and triple respectively — so the capability that distinguishes Ray 3.2 is also the one that changes its effective price the most.

On Wan 3.0 the equivalent trap is at the provider layer rather than the vendor's: when a model is priced per second per resolution, an intermediary's rate is often not Alibaba's rate. $0.05 / $0.10 / $0.20 is the number to check any quote against.

Where the routing question actually lands

Neither of these is a model we route, and it is worth saying so before anything else in this section. Wan 3.0 is reachable through Alibaba's own platforms — Model Studio and Qwen Cloud — and through third-party creative tools that integrated it after the 24 August launch. Ray 3.2 is reachable through the Luma web app and, for the first time in the Ray line, a proper developer API with pay-as-you-go and reserved-capacity tiers, plus a handful of third-party platforms and its Adobe Firefly integration.

What is available on OrcaRouter is the surrounding pipeline. A finished video is rarely the whole deliverable: there is a script, a shot list, alt text, captions in several languages, review notes, and metadata, and all of that is text-model work. Those models are in our catalogue — 200+ of them on one key, provider list price passed through with no markup added — so moving the captioning pass from an expensive model to a cheap one, or swapping the model that writes your shot lists, is a config change rather than a second integration. Given that both video models here are already a second and third vendor relationship, keeping the text layer on one key is the part of this stack that is genuinely cheap to consolidate.

Which one, and what would settle it

This is not a close call on paper, and it is not a rout either, because the two are not competing for the same job.

• Choose Wan 3.0 if you need audio from the prompt, more than 10 seconds in one pass, document or web-page grounding, a wide reference budget, 30fps, or the lowest per-second rate at 1080p. It is the better choice for anything that ends as a published clip.

• Choose Luma Ray 3.2 if you need 16-bit HDR or EXR in ACES2065-1 for a grading pipeline, frame-level control across 16 keyframes, performance tracking across eight faces, Reframe's post-hoc aspect and background changes, or source-preserving transformation of existing footage up to 20 seconds. It is the better choice for anything that ends as a shot in someone else's edit.

The thing that would change this article is not a new model — it is an arena entry. Ray 3.2 has been generally available since 9 June 2026, long enough to accrue comparisons, and it simply has not been put in front of voters. Until it is, the honest position is that Wan 3.0 has a checkable independent score and Ray 3.2 has a vendor's own report, and a comparison that pretends those are the same kind of evidence is worse than one that admits the gap.

One practical note if you are running this evaluation yourself: because Ray 3.2 outputs silent video and Wan 3.0 does not, a straight same-prompt comparison will feel lopsided against Luma in a way that has nothing to do with image quality. Score the picture first, mute both, and evaluate the audio separately — otherwise you are comparing a finished piece to a component and measuring the wrong difference.