# FLUX 3 Max Video API: Pricing, Clip Length and Generation Limits

fal's video model offers a launch discount and fast reported generation, but output pricing and real waiting time measure different things.

By Paras Madan | 2026-10-20


fal and Black Forest Labs announced FLUX 3 Max video generation on October 19. The hosted model offers text-to-video and image-to-video APIs, with 720p launch pricing of $0.0425 per output-second. The image-to-video API documents clips from 5 to 20 seconds, optional final-frame guidance and generated audio.

The [official fal announcement](https://x.com/fal/status/2112251022047412492) describes the video release. It is separate from FLUX image-model coverage: the practical output here is a video clip. Public API access is documented, but a public release of model weights was not verified.

## FLUX 3 Max pricing

The [FLUX 3 Max landing page](https://fal.ai/flux-3-max) lists a promotional 720p rate of $0.0425 per output-second, compared with a $0.17 list rate. It gives November 7 as the promotion's end date.

| 720p output duration | Promotional model cost | Listed standard cost |
| --- | --- | --- |
| 5 seconds | $0.2125 | $0.85 |
| 10 seconds | $0.425 | $1.70 |
| 20 seconds | $0.85 | $3.40 |

These examples multiply the published rate by clip duration. They are not a quote for a complete production workflow, which may involve rejected attempts, editing and other services.

The announcement posts describe the promotion's duration differently. The dated November 7 expiry on the product page is the more specific reference, but check the displayed price before submitting paid work. The discount should not be treated as a permanent rate.

The API also lists 1080p as an available resolution option, while the inspected landing-page price is for 720p. A 1080p rate was not verified, so the table should not be used to budget that setting.

## Image-to-video controls and defaults

The [image-to-video API reference](https://fal.ai/models/blackforestlabs/flux-3-max/image-to-video/api) requires a prompt and a starting image URL. It accepts PNG, JPEG or WebP, with an optional last-frame image to guide the ending.

Documented duration options span 5 to 20 seconds, with a default of 5. Resolution defaults to 720p. Audio is enabled by default, which is worth checking when the output is intended for a workflow that supplies its own sound.

The API includes low and high reasoning settings, defaulting to high. Its documentation says higher effort increases processing time. It also supports several landscape, square and portrait aspect ratios rather than forcing every clip into one format.

fal separately documents a [text-to-video endpoint](https://fal.ai/models/blackforestlabs/flux-3-max/text-to-video/api). Black Forest Labs also links a keyframes-to-video interface, but its detailed limits were not verified here. The optional final image described above belongs specifically to the inspected image-to-video API.

## How fast is FLUX 3 Max?

fal reports an average of 12.45 seconds to generate a ten-second clip in its published comparison. That is a vendor measurement under the comparison's conditions, rather than a service-level guarantee for every request.

There is another timing distinction in the API. Its returned inference timing covers denoising and excludes prompt expansion, decoding and upload. An application measuring the interval between submission and a usable video can therefore see a different duration from that field.

**Billed video duration and processing time are different quantities.** A ten-second output is priced as ten seconds of output under the published rate, regardless of whether the generation takes roughly that long or significantly longer.

For a team building a video interface, measure the complete wait under the intended resolution, reasoning setting and workload. That gives a more useful basis for progress indicators and user expectations than a single model-stage timing number.

## What the quality comparison establishes

The landing page reports a preliminary image-to-video Elo score of 1139 and describes a tie at the top based on overlapping confidence intervals. Its table also lists MiniMax H3 at 1141. That does not support a claim that FLUX 3 Max has an unambiguous numerical lead.

The independent benchmark board was not directly inspected for this article, and the model has not been tested here. Treat the result as a comparison reported by fal, rather than an independently verified ranking or a guarantee about a particular shot.

A production evaluation should include the kinds of failures that matter for the actual use case: whether the subject stays consistent, whether the requested movement is followed, and whether the ending works with the next shot. A preference score across many clips cannot answer those questions for a specific project.

## Where the launch could be useful

The combination of an explicit per-output-second price and a low reported generation time makes iterative shot development worth evaluating. A team could test several short variations before spending time assembling a longer sequence.

The relevant unit is still an accepted clip. If a lower-priced generation needs many more attempts, some of that saving disappears. Track accepted outputs alongside generation count, complete waiting time and manual editing, especially while assessing a temporary promotion.

The model's practical value will come from how quickly a usable shot reaches the edit. The API's controls and pricing make that experiment concrete; the launch benchmark alone does not settle it.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn model launches, API details and creative-tool tradeoffs into clear technical stories that developers and founders can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).