Luma AI logo

Frame-led cost sheet · checked 29 August 2026

Luma API · Ray 3.2

Direct everypaid frame.

Ray 3.2 sells a defined generation, not a guaranteed usable shot. Budget the route, duration, resolution and every reroll—then measure what the edit actually keeps.

Ray 3.2 SDR list-price matrix

Per video first.
Per second later.

Luma marks these prices approximate and based on billing tokens. T2V/I2V and V2V publish per-video totals; Reframe uses a per-second rate.

Text / image to video · SDR

One submitted generation

Duration540p720p1080p
5 sec$0.15$0.30$1.20
10 sec$0.45$0.90$3.60

Video transformation · SDR

Different route, different unit

Task540p720p1080pUnit
V2V · 5 sec$0.72$1.44$2.16Per video
V2V · 10 sec$1.08$2.16$4.32Per video
Reframe$0.06$0.12$0.36Per second

SDR output · table baseline

HDR output · five-second route

HDR + EXR · five-second route

Official approximate list prices checked August 29, 2026. The generation guide says HDR is incompatible with ten-second output. Recheck the live source before approving spend.

Retry-adjusted take ledger

Count every take.
Keep only the cut.

The default example is a calculator demonstration, not a measured Luma success rate. Replace it with your recorded attempts and retained seconds.

Ray 3.2 take ledger

Price each submitted take, then divide the total by footage that survives the edit.

Cost / usable second

720p · 5s

$0.15

At this yield, one usable finished minute projects to $9.00 in Luma generation charges.

Price / attempt
$0.30
Submitted cost
$1.80
Raw footage
30s
Usable yield
40%

Ray 3.2 SDR list-price estimate · checked 29 August 2026.

Planning estimate only. It excludes HDR/EXR multipliers, taxes, storage, sound, repair, editorial work and any account-specific terms.

Ray 3.2 direction surface

Choose the frame contract.

A first frame establishes the entrance. A last frame can define the destination. Multi-keyframe input places additional anchors at selected points on the 24 fps output grid.

Start only

0

Use a single anchor at the opening when the motion destination can remain model-directed.

Start + end

0 → 120

Pin both ends of a five-second shot. The official guide rejects this legacy pair with ten-second duration.

Multi-keyframe

0 → 60 → 120

Place multiple guide images at arbitrary output-frame positions for beat-by-beat direction.

Important validation rule

Do not send start_frame or end_frame together with keyframes. The two control surfaces are mutually exclusive and the API rejects the request.

Submit → poll → copy

The generation ID is part of the shot record.

The Agents API returns immediately with an opaque ID. Treat queued, processing, completed and failed as production states, not UI decoration.

POST /v1/generations
{
  "model": "ray-3.2",
  "type": "video",
  "prompt": "locked scene prompt",
  "aspect_ratio": "9:16",
  "video": {
    "resolution": "720p",
    "duration": "5s"
  }
}

Minimal vertical example adapted from the official guide. Add one supported frame-control strategy and validate the current schema before deployment.

  1. KF 000

    Lock the production route

    Record text-to-video, image-to-video, video-to-video or Reframe before submission; each route has a different billing unit.

  2. KF 024

    Prepare frame references

    Use a start frame, start-and-end pair or multi-keyframe array only when that control surface matches the shot design and selected duration.

  3. KF 048

    Create the generation

    Submit Ray 3.2 with an explicit aspect ratio, resolution and duration, then preserve the returned opaque generation ID.

  4. KF 072

    Poll the task state

    Wait until the generation becomes completed or failed; log failures and retries rather than counting only successful downloads.

  5. KF 096

    Copy the output promptly

    Download the MP4 from the completed output URL into controlled storage because the API response uses a presigned URL.

  6. KF 120

    Price the accepted edit

    Sum every paid attempt and divide by the seconds retained after continuity, performance and editorial review.

Same locked vertical-drama pilot

Anchor the beat.
Compare the take.

These are the same three controlled prompts used in the Runway and Kling guides. Keep ratio, duration and reference assets fixed before judging a route.

KF 0 → 120

Shot 01

A / dialogue two-shot

Identity, lip sync and eyeline continuity

Vertical 9:16 medium two-shot inside a late-night convenience store. Mara, a woman in a charcoal coat with a small silver crescent pin, confronts Elias, a tired man in a pale blue work shirt. Keep both faces visible and preserve their exact wardrobe and relative positions. Mara says, ‘You knew the train was empty.’ Elias answers, ‘I knew who was waiting.’ Natural fluorescent ambience, restrained performances, no camera cut, subtle push-in, realistic synchronized dialogue and room tone.

KF 0 → 60 → 120

Shot 02

B / reaction close-up

Face, wardrobe and emotional continuity

Vertical 9:16 close-up of Mara in the same late-night convenience store, wearing the same charcoal coat and silver crescent pin. She processes Elias’s answer without speaking: suspicion shifts into recognition, then controlled fear. Match the fluorescent direction and background geography from the prior two-shot. Locked camera, shallow depth of field, realistic micro-expression, no new props, no text, continuous room tone.

KF 0 → 48 → 120

Shot 03

C / moving reveal

Motion, geography and prop continuity

Vertical 9:16 moving reveal in the same convenience store. Begin behind Elias’s right shoulder, track slowly past him and reveal a red suitcase beneath the counter behind Mara. Preserve both characters’ faces, wardrobe, store geography and fluorescent lighting. Elias remains foreground right; Mara remains behind the counter. The suitcase is the only new reveal. Controlled dolly movement, realistic reflections, no dialogue, tense refrigerator hum, no cuts.

The frame numbers illustrate a five-second, 24 fps planning grid. This is a test protocol, not a published Luma performance result; footage and retry data remain pending a legally obtained recorded run.

Luma pricing FAQ

Before frame zero.

How much does the Luma Ray 3.2 API cost?

For standard-dynamic-range text-to-video or image-to-video, the official Luma API table checked August 29, 2026 lists a five-second generation at about $0.15 for 540p, $0.30 for 720p and $1.20 for 1080p. Ten-second totals are about $0.45, $0.90 and $3.60 respectively.

Is Luma priced per second or per video?

It depends on the operation. The current table prices T2V/I2V and V2V as totals per video for published duration and resolution combinations. Reframe is priced per source second. The calculator preserves that distinction.

How much is a five-second 720p Luma API video?

The current Ray 3.2 SDR T2V/I2V table lists approximately $0.30 for a five-second 720p generation. Six paid attempts therefore total approximately $1.80 before taxes, post-production or account-specific terms.

Does Luma Ray 3.2 support vertical 9:16 video?

Yes. The official generation guide lists 9:16 among the accepted aspect ratios. Set it explicitly and verify the returned frame, because aspect ratio alone does not prove character, prop or composition continuity.

Can Ray 3.2 use a first frame and last frame?

Yes for a five-second generation. The guide accepts video.start_frame, video.end_frame or both. It also offers multi-keyframe image-to-video for anchors at selected output-frame positions. The legacy start/end pair is not supported with a ten-second duration.

How does the Luma API return a generated video?

POST /v1/generations returns an ID immediately. Poll GET /v1/generations/{id} until the state is completed or failed. A completed output contains a presigned MP4 URL, which should be copied into your own controlled storage.

Are ReelPulse's Luma retry and acceptance rates available?

Not yet. This page publishes official prices, capabilities, a cost model and a locked prompt protocol. Sample footage, attempt counts and acceptance data remain pending until a legally obtained recorded test run is completed.