FLUX 3 Text to Video: Features, Real Costs and How It Compares

Black Forest Labs spent two years being the company you went to for images. FLUX 3 is the first time they have shipped a video model, and the interesting part is not simply that it makes video. It is that image, video, audio and robot action prediction all come out of one jointly trained set of weights, rather than a video decoder bolted onto an image model after the fact.

I spent an afternoon running flux 3 text to video on Segmind to answer the questions I actually had: what does a clip cost, what do you get for the money, and how does it stack up against Seedance 2.5, Grok Imagine Video and Gemini Omni Flash. This post has the parameter space, four real clips with the exact amount each one billed, the arithmetic on the draft workflow, a spec comparison table, and the Python to call it.

What FLUX 3 text to video actually does

The endpoint is flux-3-text-to-video and the required parameter list is one item long: prompt. Everything else has a default. That is worth saying plainly, because plenty of video APIs make you specify six things before you can see anything at all.

The parameter space, read from the model's own spec rather than guessed at:

• duration: any whole number of seconds from 5 to 20. Not an enum of three fixed options, which is what Gemini Omni Flash and Veo 3.1 give you. If you need a 13 second cut, you can ask for 13 seconds.
• resolution: hd or fhd.
• aspect_ratio: auto, 21:9, 2:1, 16:9, 4:3, 1:1, 3:4, 9:16.
• draft: a boolean that switches the job to a cheap preview tier.
• generate_audio: defaults to true. You opt out, not in.

That last one deserves a flag. Audio being on by default means the first clip you generate has sound whether you planned for it or not, and it means every cost quoted in this post is a cost with audio included. All four of my clips came back carrying a real AAC track at 44.1 kHz.

There are four FLUX 3 endpoints on Segmind, not one: flux-3-text-to-video, flux-3-image-to-video, flux-3-extend-video, and flux-3-draft-enhance. The last one is the interesting one and I will come back to it.

HD and FHD do not mean what I assumed

I assumed hd meant a 1280 pixel long edge. It does not. HD is a pixel budget of roughly 0.9 megapixels with both edges snapped to a multiple of 32, so the frame size moves with the aspect ratio you ask for. Reading the container atoms straight off the files I generated:

• 16:9 came back 1280x704, which is 901,120 pixels.
• 21:9 came back 1440x608, which is 875,520 pixels.
• FHD came back 1920x1088.

Nothing there is exactly 720p and nothing is exactly 1080p. If you have a pipeline that assumes a 1280 pixel long edge, it breaks the first time somebody requests a wide aspect ratio. Worth knowing before you write the ingest code rather than after.

What FLUX 3 text to video actually costs

Every FLUX 3 page lists the same pricing line: GPU Time, $0.0072 per GPU second. That describes the meter, and you cannot plan a budget with it, because you have no way of knowing how many GPU seconds a render will burn before you run it.

What actually happens is far easier to work with. You are billed a flat rate per second of finished video, set by the quality tier. I captured the x-cost response header on every fire, so the table below is billed amounts rather than estimates.

Three things fall straight out of that table.

The rate is exactly linear in duration. The 5 second HD clip billed $1.2155 and the 10 second HD clip billed $2.4310, which is precisely double. There is no fixed per-request component hiding in there, so seconds multiplied by tier rate gives you your bill to the cent before you fire anything.

Aspect ratio is free. The 16:9 and the 21:9 clips both billed at $0.2431 per second, despite 16:9 carrying 901,120 pixels per frame against 21:9's 875,520. You are charged for duration and tier, not for pixels. Pick the framing your edit needs and ignore the cost column.

Render time tells you nothing about price. The 346 second render cost $4.15 and the 62 second render cost $1.22, but the 96 second render cost $0.86. Wall clock and bill are unrelated, so set a generous client timeout and do not try to infer cost from how long something took.

The draft workflow, and the arithmetic that justifies it

This is the feature I would build a pipeline around. Set draft: true and you get a cheap pass. Read X-Draft-Cache-Url off the response headers, hand it to flux-3-draft-enhance, and the model re-renders that exact cached generation at full quality, keeping the same seed and the same motion.

The first thing to understand is that draft is a quality tier, not a resolution drop. My draft came back at 1280x704 with a full audio track: the same frame size as the full-quality HD pass, for a third of the price. You are buying fewer sampling steps, not fewer pixels.

The second thing is that drafting is not free money, and it is worth being precise about where the saving actually comes from. A finished FHD clip by the draft route cost me $5.005 all in: $0.858 for the preview plus $4.147 for the enhance. On a clip you were always going to keep, the preview is a surcharge.

Where it wins is on the takes you throw away. Rejecting a draft costs $0.858. Rejecting a finished FHD render costs $4.147. That is 4.8 times cheaper per rejected take, and since nobody gets the shot on the first try, that is the number that governs a real production budget. Generate ten candidates and enhance one and you have spent $12.73 instead of $41.47.

One gotcha that cost me a fire to learn: flux-3-draft-enhance accepts only draft_cache and resolution. There is no duration parameter, because the length is baked into the cached draft. If you want a 12 second final, you must have drafted at 12 seconds.

Use case 1: marketing agencies, and the draft-first loop

An agency turning out product video variants has one economic problem: most of what it generates gets discarded, and you cannot tell which take works until you see it. So I ran the same product prompt three ways, holding everything else constant at 10 seconds and 16:9.

That preview cost $0.858 and it is not a low-resolution proxy: same 1280x704, same audio track, 96 seconds to render. This is the artifact you circulate internally to decide whether the idea works at all.

Here is the same prompt at full quality, billed $2.4310:

And here is the draft above pushed through flux-3-draft-enhance to Full HD, which re-renders the cached generation at 1920x1088 while holding the same seed and motion:

One measurement I did not expect: the draft file is larger on disk than the full-quality render, 3.59 MB against 2.29 MB for identical dimensions and duration. Fewer sampling steps plausibly leaves more high-frequency noise for the encoder to carry, which would explain it, though that is an inference from file size rather than something I can prove without watching them frame by frame.

Use case 2: film and previs, at 21:9 with dialogue

For anything cinematic, two parameters matter more than the rest: the aspect ratio, and whether the audio can carry a scripted line. Previs that cannot hold a spoken beat is a moodboard that moves. FLUX 3 offers a true 21:9 and lip-synced dialogue, so I asked for both at once.

The container came back at 1440x608 with an AAC track, in 62 seconds, for $1.2155. That is the cheapest and fastest fire of the four, and it is the one carrying dialogue, which is a reasonable summary of why the per-second pricing model is pleasant to work with: a short cinematic beat is genuinely cheap, and the wide framing costs nothing extra.

For a previs workflow, the practical shape is to draft the whole sequence at 21:9, cut it, and only enhance the shots that survive the edit. At $0.0858 per second you can previs a two minute sequence for about $10.

Use case 3: production houses and MCNs, vertical at volume

Vertical is where the volume is, and volume is where per-clip cost stops being a rounding error. FLUX 3 exposes 9:16 and 3:4 directly, and since aspect ratio does not affect the rate, the arithmetic for a channel network is simple.

A network shipping 200 short vertical clips a month at 10 seconds each is buying 2,000 seconds of video. At the draft rate that is $171.60. At full-quality HD it is $486.20. At Full HD it is $829.40. Those three numbers are the entire strategic decision, and the draft-first workflow lets you sit at the first number for everything that is still being decided and pay the third only for the handful that ship.

The parameter that makes this practical is per-second duration control. Most short-form platforms want a specific length, and being able to request exactly 7 or exactly 13 seconds rather than generating 15 and trimming means you are not paying for footage you cut away.

Calling it from Python

The response body is the MP4 itself, not a JSON envelope pointing at a file you then have to fetch. Write response.content straight to disk.

import requests

resp = requests.post(
    "https://api.segmind.com/v1/flux-3-text-to-video",
    headers={"x-api-key": "YOUR_API_KEY"},
    json={
        "prompt": "Cinematic macro product shot: a frosted glass serum bottle on wet black slate, morning light raking in from the left, water beading and running down the glass. Soft ambient room tone, no music.",
        "duration": 10,
        "resolution": "hd",
        "aspect_ratio": "16:9",
        "draft": True,
        "generate_audio": True,
    },
    timeout=900,
)

open("preview.mp4", "wb").write(resp.content)
print("billed:", resp.headers.get("x-cost"))
print("draft cache:", resp.headers.get("x-draft-cache-url"))

Then the enhance step, which is the whole point of drafting:

cache_url = resp.headers["x-draft-cache-url"]

final = requests.post(
    "https://api.segmind.com/v1/flux-3-draft-enhance",
    headers={"x-api-key": "YOUR_API_KEY"},
    json={"draft_cache": cache_url, "resolution": "fhd"},
    timeout=900,
)
open("final.mp4", "wb").write(final.content)

Two operational notes. Read x-cost on every call, because it is the only reliable record of what a request actually billed. And set the client timeout high: my fires ran from 62 to 346 seconds, and a render continues on the server after your client gives up, so a short timeout loses you the file while still costing you the money.

How FLUX 3 compares

Spec comparison only, no output samples. I read every row below off each model's own spec page on the same day, because vendor comparison tables have a habit of going stale in whichever direction flatters the vendor.

A few things that table hides. Seedance 2.5 is metered in tokens rather than seconds, so putting it in a dollars-per-second column requires a conversion that itself depends on resolution. Grok Imagine Video advertises a 1080p option in its resolution enum that I would test before depending on it. And Gemini Omni Flash and Veo 3.1 accept only a few fixed durations, which sounds minor until you need a 7 second cut.

LTX 2.5 is not available on Segmind, so I left it out rather than quote numbers from elsewhere and present them as comparable. The only LTX endpoint on the platform is the original LTX-Video, a 768x512 model with no audio track, which is not the same class of thing.

Honest assessment

What it does well: the draft tier is the most genuinely useful cost feature I have used on a video API, precisely because it is a quality tier rather than a resolution drop, and the cache handle means the cheap pass is not throwaway work. Per-second duration control and free-by-default audio remove two more reasons I would normally reach for a different model.

Where it falls short: 20 seconds is the ceiling and Seedance 2.5 does 30. Latency swung from 62 to 346 seconds across four fires with no relationship to duration or cost, so this is not a model to put behind a user-facing spinner. And I would want to see FHD compared against a native 1080p render before treating the top tier as a true 1080p deliverable.

FAQ

What is FLUX 3 text to video?

FLUX 3 text to video is Black Forest Labs' first video model, served on Segmind at flux-3-text-to-video. You send a prompt and get back an MP4 of 5 to 20 seconds with a synchronised audio track, in HD or Full HD, across eight aspect ratios.

How long can a FLUX 3 clip be?

Up to 20 seconds, and you can ask for any whole number of seconds from 5 upward. That is unusual: most video APIs only accept a handful of fixed durations, so a 13 second cut means generating 15 and trimming.

Does FLUX 3 generate audio?

Yes, and it is on by default, so you opt out rather than in. All four clips I generated came back with a real AAC track at 44.1 kHz, and every price quoted in this post is therefore a with-audio price.

What is the draft workflow in FLUX 3 text to video?

Set draft: true for a cheap preview pass. The response carries an X-Draft-Cache-Url header. Feed that to flux-3-draft-enhance and it re-renders the same seed and the same motion at full quality, so you only pay the expensive rate for takes you have already approved.

Is FLUX 3 cheaper than Seedance 2.5?

At the draft tier, yes, and by a wide margin. At full quality the two are much closer, and Seedance 2.5 goes to 30 seconds and 1080p where FLUX 3 stops at 20 seconds. The right answer depends on whether you are generating volume or finals.

How do I call FLUX 3 text to video from Python?

POST to https://api.segmind.com/v1/flux-3-text-to-video with your x-api-key header and a JSON body containing at minimum a prompt. The response body is the MP4 itself, not a JSON envelope with a URL, so write response.content straight to a file.

Conclusion

FLUX 3 text to video earns its place for volume work: draft everything, enhance the keepers, and let the free audio do a job you would otherwise pay a second vendor for. If you need 30 second takes, look at Seedance 2.5. If you need clips under 10 seconds at the lowest price per second, Grok Imagine Video is cheaper.

Every call in this post runs on the FLUX 3 text to video model page, and the enhance step is at flux-3-draft-enhance. No setup, no GPU, one POST.