FLUX 3 Video Review: Benchmarks, API Pricing, and Test

👁 119 reads

Continuously updated. Status as of October 11, 2026.

Update, October 20, 2026: FLUX 3 Max, the faster FLUX 3 Video variant post-trained by fal, delivered a 10-second clip in 20-25 seconds in my test. Details, prices and clips: FLUX 3 Max: 10 s of AI Video in 12 s? My Test, Prices, Clips.

FLUX 3 Video is the first video model from Germany-based Black Forest Labs, the creators of FLUX image models. It generates 5 to 20 second clips with integrated audio, costs $0.17 per second in 720p via the official API (or $0.085 on fal.ai), and ranks 6th on the Artificial Analysis text-to-video leaderboard with an Elo score of 1128. I generated 8 clips with it: it reproduced non-English speech verbatim, rendered tricky accented typography without flaws, and scored an average of 8.9/10 across five scenes in my Video Arena benchmark - the highest score among all models I have evaluated there.

What Is FLUX 3 Video?

Black Forest Labs, based in Freiburg im Breisgau, Germany, was founded by the original creators of Stable Diffusion. Prior to summer 2026, the company was known for its image generation models, FLUX.1 and FLUX.2. FLUX 3 is their first foundational model trained from scratch natively on images, video, audio, and motion prediction simultaneously. The lab distributes it across several editions:

• FLUX 3 Video - Video generation and editing with audio via API. This article covers this model.
• FLUX 3 Image - Still images and editing, launched October 1, 2026. Covered in my FLUX 3 Image guide.
• FLUX 3 Action - Motion prediction for robotics, developed with partners.
• FLUX 3 Dev - An announced open-weights edition, release date pending.

In its August 4 announcement, BFL emphasized that the model is designed to step beyond sterile “cinematic” aesthetics - generating clips with raw, candid, comedic, or nostalgic visual signatures. It handles in-shot camera transitions, renders legible signage as a natural part of the scene, and generates multi-character lip-synced speech in numerous languages. I previously discussed early access findings in August.

Artificial Analysis Leaderboard Ranking

Data retrieved from the Artificial Analysis AA-Video-T2V v2.0 leaderboard on October 11, 2026. The $17.40 price matches the 1080p tier from BFL’s official rate sheet ($0.29 x 60 s). The confidence intervals for FLUX 3 (1119-1137) and MiniMax H3 Max (1120-1140) overlap, making ranks 4 through 6 a statistical tie.

Scale starts at 1,000 points (Veo 3.1 falls below base scale).

FLUX 3 Video Specifications

While some regional languages are not explicitly listed by name, BFL notes broader language support under “and more”, which my tests verified. Sources: BFL August 4 blog post, documentation, and fal.ai input schema.

FLUX 3 Video API Pricing

Official Black Forest Labs API rates (1 credit = $0.01), verified October 11, 2026. Third-party provider rates:

• fal.ai - $0.085/s for text, image, and keyframe video; draft mode at $0.03/s (fal.ai rates as of Oct 11, exactly half of BFL’s direct pricing). fal.ai has not stated how long this promotional rate will last.
• OpenRouter - $0.17/s at 720p and $0.29/s at 1080p, matching direct BFL pricing. Draft mode is unavailable here.
• Vercel AI Gateway - Starts at draft mode tier, $0.06/s.

Each of my 6 runs via OpenRouter billed exactly $0.85 (verified in usage.cost), while 2 runs on fal.ai billed $0.425 each. For a wider industry comparison, see my overview of how much AI video generation costs per second.

Hands-on Test: FLUX 3 in the Video Arena

On October 11, I subjected FLUX 3 to the exact same five test prompts that six other models received on October 8 in my Video Arena: 5-second duration, 16:9 aspect ratio, single-pass generation with zero cherry-picking. Note: FLUX 3 lacks a 480p mode, so its clips ran at 720p, whereas Wan and MiniMax were evaluated at 480p. All clips were graded on the same 0-10 scale.

The 0.1-point advantage over Wan 3.0 is effectively a tie, and Wan achieved its result in 480p at one-fourth the cost. However, FLUX 3 became only the third model in the Arena (alongside Wan 3.0 and H3 Max) to handle complex diacritics flawlessly, and one of the rare models to articulate multi-syllabic foreign phonetic clusters without slurring - previously matched only by H3 Max, Gemini Omni Flash 1.1, and Grok Imagine 1.5. You can inspect all prompt inputs and outputs in my roundup: What FLUX 3 Video Can Do: 8 Real-World Examples.

Multilingual Text and Speech Performance

I specifically tested the model on difficult foreign phonetic strings and non-Latin typography using Polish as a benchmark stress test:

• Speech output:Generated two distinct speech clips with zero phonetic drift. Both a classic tongue-twister (translated from the Polish original: “In Szczebrzeszyn a beetle buzzes in the reed…”) and standard commercial voice-over lines were transcribed word-for-word by Whisper.
• On-screen typography:Accented characters such as Polish dotted Z, stroke L, O with acute, and Z with acute rendered cleanly on the first pass without malformed glyphs.
• Prompting recommendations: I recommend writing structural scene descriptions, camera framing, and lighting instructions in English, while inserting non-English dialogue directly inside quotation marks after explicit cues like “says in Polish” or “says in Spanish”.

How to Get Started with FLUX 3 Video

1. Choose an API provider: direct Black Forest Labs API ($0.01 per credit), fal.ai (most economical at $0.085/s), or OpenRouter (convenient multi-model routing).
2. Set up an account, top up your balance with a few dollars, and generate your API key.
3. Prototype using draft mode (available on BFL and fal.ai): a 10-second draft costs roughly $0.30-$0.60, letting you lock down framing and pacing before committing to full refinement.
4. Structure your prompt clearly: describe scene layout, camera moves, lighting cues, and audio expectations.
5. Dispatch the API request and poll the job ID - in my tests, 5-second 720p generations completed in 47 to 98 seconds.

For a production-ready integration, see my practical code guide: How to Generate Video with FLUX 3 Video via API.

FLUX 3 Video vs. Competing Models

For an in-depth breakdown, read my head-to-head analysis: FLUX 3 Video vs. Wan 3.0 vs. Seedance 2.5. For the broader landscape, consult my guides on top AI video generators, my full AI video model leaderboard, and the latest on Vidu Q4 for image-to-video.

Who Should Use FLUX 3 Video?

• Commercial creators requiring reliable speech and typography - Dialogues and on-screen signage render accurately without requiring multiple rerolls.
• Developers building automated media pipelines - Supported out of the box across three major API aggregators, with discounted execution available via fal.ai.
• Iterative creators - The low-cost draft mode ($0.03-$0.06/s) makes visual pre-visualization budget-friendly before rendering final passes.
• Not ideal for low-cost mass production - If you only need standard-definition 480p output, Wan 3.0 and MiniMax H3 are substantially cheaper per minute.
• Not for non-technical users looking for a consumer UI - BFL does not currently provide a dedicated first-party consumer web interface; access requires third-party frontends or API keys.

Frequently Asked Questions

How much does FLUX 3 Video cost?

Through the official BFL API, pricing is $0.17 per second for 720p and $0.29 per second for 1080p, with integrated audio included at no surcharge. On fal.ai, standard generation costs $0.085 per second. A 5-second 720p clip costs $0.85 on BFL or OpenRouter, and $0.425 on fal.ai.

Is FLUX 3 Video available for free?

There is no official free plan. However, Vercel AI Gateway provides $5 in monthly platform credits on free-tier developer accounts, which is sufficient to render several draft-mode tests.

Can FLUX 3 Video speak multiple languages?

Yes. Beyond its officially listed core languages, it handles multilingual speech and complex regional accents reliably. In my hands-on benchmark, Whisper verified that test Polish speech clips generated word-for-word without dropouts.

What is the maximum clip duration in FLUX 3?

A single generation pass yields between 5 and 20 seconds. Longer continuous sequences can be created using the clip extension endpoint, which uses up to 4 seconds of context from the preceding shot to generate an additional 5 to 15 seconds.

Does FLUX 3 Video support native 4K output?

The core generative architecture produces frames at 720p, which are upscaled to 1080p, QHD, or 4K via integrated spatial upscalers. 4K generation is billed at $0.80 per second.

How does FLUX 3 Video differ from FLUX 3 Image?

They share underlying foundational architecture, but are distinct endpoints: FLUX 3 Image handles still image synthesis up to 4K resolution (launched October 1, 2026), while FLUX 3 Video focuses on motion and audio generation (launched August 4, 2026). For images, read my dedicated FLUX 3 Image guide.

Are the FLUX 3 model weights open source?

No. While Black Forest Labs has announced plans for an open-weight FLUX 3 Dev release, no release schedule has been shared.

Related Guides

• What FLUX 3 Video Can Do: 8 Real Prompts and Results
• FLUX 3 Video vs. Wan 3.0 vs. Seedance 2.5
• How to Generate Video with FLUX 3 Video via API: Code and Tutorial
• FLUX 3 Image: 4K Synthesis and Image Editing Guide