FLUX 3 Max: The Frontier Video Model That Renders 10 Seconds in 12

What FLUX 3 Max is

Black Forest Labs released FLUX 3 in the summer of 2026 as its first model trained jointly on images, video and audio, and described it as a base designed for customization. FLUX 3 Max is the first big example of that: fal's research team took the FLUX 3 foundation model and post-trained it for visual fidelity, motion coherence, prompt adherence and image-to-video quality, then optimized it on fal's inference engine to cut latency.

The result is a narrower, faster model than FLUX 3 itself. It focuses on two jobs, text-to-video and image-to-video (with an optional last frame), and drops FLUX 3's draft, keyframe and video-extension endpoints. In exchange it is about ten times faster than the model it came from and scores 50 Elo points higher.

"Tied for #1": what the benchmark says

The quality claim comes from the Artificial Analysis Video Arena, where human raters compare two videos generated from the same input and pick the better one. On the image-to-video board (AA-Video-I2V v1.1), preliminary results at launch looked like this:

Read it as a three-way tie at the top, not a clear win. MiniMax H3 is two points ahead by point estimate, the confidence intervals overlap (FLUX 3 Max's 95% interval spans roughly 1,123 to 1,153, which puts its possible rank anywhere from first to fifth), and the results were still preliminary at launch. What the board does show clearly is the jump from FLUX 3 (1,089) to FLUX 3 Max (1,139) after post-training.

Speed is the real headline

Where FLUX 3 Max stands apart is latency. fal published the average time to generate a 10-second, 16:9 clip at 720p or 768p:

That is about 22 times faster than Seedance 2.5 and 10 times faster than FLUX 3, at the same or better quality. These are fal's own measurements, not an independent test, but they match what we saw: in our first test call, a 5-second 720p clip with reasoning_level set to low came back in about 12 seconds end to end, queue included.

Speed changes how you use a frontier model. At a few minutes per clip, you write a prompt, wait, and hope. At ten seconds, you can iterate on a prompt the way you would on an image, generate several variants for a user to choose from, or put generation inside an interactive product without a progress bar.

Specs

How to use FLUX 3 Max on VideoGenAPI

FLUX 3 Max uses the same endpoint and API key as every other model on VideoGenAPI. Text-to-video:

curl -X POST "https://videogenapi.com/api/v1/generate" \
  -H "Authorization: Bearer $VIDEOGEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "flux-3-max",
    "prompt": "A red panda walks along a mossy log in a sunlit forest, pausing to sniff the air. Ambient birdsong.",
    "duration": 10,
    "resolution": "720p",
    "aspect_ratio": "16:9"
  }'

For image-to-video, add the first frame and, optionally, the last one. The image is where FLUX 3 Max ranks best:

import os, time, requests

API = "https://videogenapi.com/api/v1"
HEADERS = {"Authorization": f"Bearer {os.environ['VIDEOGEN_API_KEY']}"}

job = requests.post(f"{API}/generate", headers=HEADERS, json={
    "model": "flux-3-max",
    "prompt": "The bottle turns slowly on its base, condensation running down the glass, soft studio light",
    "image_url": "https://example.com/product.jpg",
    "duration": 8,
    "reasoning_level": "low",
}).json()

while True:
    status = requests.get(f"{API}/status/{job['generation_id']}", headers=HEADERS).json()
    if status["status"] in ("completed", "failed"):
        break
    time.sleep(3)

print(status.get("video_url") or status.get("error"))

Optional fields: end_image_url for a last frame, generate_audio to turn sound off, seed to reproduce a result. The full parameter list is on the FLUX 3 API page and in the generate reference.

Pricing

On fal, FLUX 3 Max costs $0.17 per second at 720p and $0.29 at 1080p, with a 75% launch discount until November 7, 2026. On VideoGenAPI it is billed from your balance at $0.25 per second at 720p and $0.40 at 1080p:

That is more than fal's list price, and we say so plainly: if FLUX 3 Max is the only model you need, calling fal directly is cheaper. VideoGenAPI makes sense when FLUX 3 Max is one model among several, on one key and one bill, next to models included in the monthly plans such as Kling 3 and MiniMax H3. Failed FLUX 3 Max generations are refunded to your balance automatically. For the wider market, see our AI video API pricing comparison.

When to pick FLUX 3 Max

• Animating an existing image: product shots, illustrations, key art. Image-to-video is where it ranks highest.
• Anything interactive: previews, variant pickers, user-facing generation where waiting minutes is not acceptable.
• Clips up to 20 seconds in one generation, with sound.
• Prompt iteration: at around ten seconds per render, set reasoning_level to low while exploring, then switch to high for the final.

When cost per clip matters more than speed, models included in a VideoGenAPI plan such as MiniMax H3 or Kling 3 are much cheaper at volume. For the earlier FLUX 3 release and how it compares with Veo 3.1 and Kling 3, read FLUX 3 vs Veo 3.1 vs Kling 3.

Frequently asked questions

Sources