FLUX 3 is Black Forest Labs’ first video model: it makes clips of up to 20 seconds with sound, through a paid API, and you cannot download it. The company best known for FLUX image models announced it on 23 July 2026 and opened video generation on 4 August.
- One model for picture and sound. Speech, effects and ambience are generated with the frames, and audio costs nothing extra.
- Priced by the second. Four resolutions from 720p to 4K, plus a cheap draft mode for trying ideas before you pay full price.
- Not the cheapest. Next to the rates on this site’s comparison page, it costs more per second than Veo 3.1 Lite, Omni or Grok at 720p, and more than Veo 3.1 at 4K.
- No open weights yet. An open-weight version is promised, not released. The open FLUX.2 [klein] models are a different thing.
- You can use the output commercially, but the terms also let Black Forest Labs train on what you send and get back.
Every price here was correct when this site last checked it, between 16 Sep and 23 Sep 2026. Prices change often and this site no longer updates them, so check the vendor’s own page before you rely on one: Black Forest Labs pricing.
What FLUX 3 does
Black Forest Labs (BFL) describes it in one line: “FLUX 3 is our first video model: one model trained across image, video, and audio, generating video with synchronized sound.” The same entry gives clips of up to 20 seconds at 1080p and 24 frames a second from one request.Black Forest Labs, Release notes, 4 Aug 2026 entry, read at source 22 Sep 2026 and unchanged on 23 Sep 2026: “Up to 20 seconds at FHD (1920 × 1088 for 16:9), 24 fps, in a single request”.
There are four ways to start a clip: from text, from one or more still images pinned as keyframes, by continuing an existing clip, or by re-rendering a draft at full quality. Continuing a clip is capped lower, at 15 seconds since 17 August. On 10 September, 2K and 4K output were added.Black Forest Labs, Release notes, read at source 22 Sep 2026 and re-read 23 Sep 2026: 4 Aug 2026, “Four modes on one endpoint”; 17 Aug 2026, “The maximum duration for video continuation” requests “is now 15 seconds”; 10 Sep 2026, “FLUX 3 Video — 2K and 4K”.
The higher resolutions are not a different shot. Black Forest Labs says every class above 720p runs the same generation and only the video upsampler’s canvas changes, “so the shot is identical across all four and only the delivered size differs.” Audio is on by default and can be switched off.Black Forest Labs, Video (FLUX 3 documentation), read at source 22 Sep 2026: generate_audio “On by default.”; re-read 23 Sep 2026: “only the delivered size differs”.
Its own launch post names hd as 720p and fhd as 1080p.Black Forest Labs, FLUX 3 Video, Part 1: Generation, 4 Aug 2026, read at source 22 Sep 2026: “We are releasing our model at HD (720p) and Full HD (1080p) resolutions”.
Choose the resolution for where the clip will be shown, not to get a better take. Every tier is the same generation, delivered at a different size.
What “preview” means for you
BFL has used three labels. The July announcement said “FLUX 3 is now available in Early Access.” The 4 August release note says “Available now as a preview.” The blog post that same day says the video model “is generally available via the BFL API and select partners.”Black Forest Labs, FLUX 3 - Real World Models, 23 Jul 2026; Release notes, 4 Aug 2026 entry; FLUX 3 Video, Part 1: Generation, 4 Aug 2026. All read at source 22 Sep 2026. Anyone can pay to use it, but the documentation still calls it a preview.Black Forest Labs, FLUX 3 model page, read at source 22 Sep 2026: “available through the Black Forest Labs dashboard and API on a pay-as-you-go basis”.
In practice that means the model can change under you. The API’s version setting defaults to latest, which “always tracks the current release”, and the API terms say the company “may modify, change, update, and/or enhance the FLUX API and/or the FLUX AI Models” at any time.Black Forest Labs, Video, read at source 22 Sep 2026 · FLUX API Service Terms, last revised 4 Aug 2026, read at source 22 Sep 2026.
BFL’s own quickstart gives two more working limits: generation takes “several minutes per generation right now”, and you should “download every result within 2 hours.”Black Forest Labs, Video quickstart, read at source 22 Sep 2026. The same page’s settings table lists resolutions of 480p and 720p, which does not match the API reference, so treat it as partly out of date. The API terms say the same about storage in general: the company “does not guarantee long-term storage of Video Outputs”.Black Forest Labs, FLUX API Service Terms, last revised 4 Aug 2026, read at source 22 Sep 2026.
Save every clip as soon as it finishes, and note the date you made it. The model you are using may not be the same one next month.
What a clip costs
Black Forest Labs charges per second of output. Text-to-video and image-to-video cost $0.17 a second at hd, $0.29 at fhd, $0.40 at qhd and $0.80 at uhd. A draft costs $0.06 a second. Continuing a clip costs more: $0.41 a second at hd up to $0.95 at uhd. Its model page adds: “Audio is included at no extra charge.”Black Forest Labs, API Pricing, read at source 22 Sep 2026 and unchanged on 23 Sep 2026: “Full-render rates are per second of output, by resolution class” · FLUX 3 model page, read at source 22 Sep 2026.
Multiplying rate by length gives these prices for one text or image clip:
- Draft, 720p: 5 s × $0.06 = $0.30; 10 s × $0.06 = $0.60.Reasoning, September 2026 — published rate × seconds.
- hd, 720p: 5 s × $0.17 = $0.85; 10 s × $0.17 = $1.70.Reasoning, September 2026 — published rate × seconds.
- fhd, 1080p: 5 s × $0.29 = $1.45; 10 s × $0.29 = $2.90.Reasoning, September 2026 — published rate × seconds.
- qhd, 2K: 5 s × $0.40 = $2.00; 10 s × $0.40 = $4.00.Reasoning, September 2026 — published rate × seconds.
- uhd, 4K: 5 s × $0.80 = $4.00; 10 s × $0.80 = $8.00.Reasoning, September 2026 — published rate × seconds.
BFL’s own figure agrees: “Five seconds of standard text-to-video in HD costs $0.85.” At the 20-second maximum, one 4K clip is $16.00.Black Forest Labs, FLUX 3 model page, read at source 22 Sep 2026: “Five seconds of standard text-to-video in HD costs $0.85.”Reasoning, September 2026 — every other total above is the published per-second rate multiplied by the clip length; Black Forest Labs does not print them.
Drafts are what bring the cost down. A draft “returns a fast preview of your prompt”, and when it is right, “FLUX 3 will render the same video at full quality”, with the upgrade “priced at the corresponding standard FHD rate.” Three 10-second drafts and one 1080p commit cost 3 × $0.60 + $2.90 = $4.70. Four full 1080p attempts cost 4 × $2.90 = $11.60.Black Forest Labs, FLUX 3 model page, read at source 22 Sep 2026: “priced at the corresponding standard FHD rate”.Reasoning, September 2026 — assumes one committed 1080p clip per three drafts; your ratio will differ.
Try ideas as drafts and only pay the full rate for the one you keep. At 4K, one full clip costs more than thirteen drafts of the same length.
How it compares on price
The other rates below are the ones on this site’s comparison of Veo 3.1, Gemini Omni and Grok Imagine Video. They were not re-read for this page. Per second: Veo 3.1 runs from $0.05 at 720p (Lite) to $0.60 at 4K (Standard); Gemini Omni Flash is about $0.10 at 720p; Grok Imagine Video is $0.08. For a 5-second 16:9 clip, BytePlus works out Seedance 2.5 at $0.231 a second at 720p and $0.569 at 1080p.This site, Veo 3.1 vs Gemini Omni vs Grok Imagine Video, checked 22 Sep 2026, citing Google’s Gemini Developer API pricing (read 16 Sep 2026), xAI’s pricing (read 16 Sep 2026) and BytePlus ModelArk Pricing (read 22 Sep 2026). Its table: “$0.05–$0.60 a second” (Veo 3.1), “about $0.10 a second at 720p” (Omni), “$0.08 a second” (Grok).
Ten seconds at those rates comes to $0.50 to $6.00 for Veo 3.1, about $1.00 for Omni at 720p and $0.80 for Grok. Five seconds of Seedance 2.5 comes to about $1.16 at 720p and $2.85 at 1080p. FLUX 3 at 720p costs $1.70 for ten seconds, or $0.85 for five.Reasoning, September 2026 — published per-second rates multiplied by clip length. Seedance is shown at 5 seconds only because BytePlus’s per-second figures are for a 5-second clip and it bills by token, so they may not scale in a straight line.
So at 720p FLUX 3 costs more than Veo 3.1 Lite, Omni and Grok, and less than Seedance 2.5. At 4K, $0.80 a second is more than Veo 3.1 Standard’s $0.60. Its 20-second maximum is longer than Grok’s 15 or Omni’s 10, and audio is included. None of these prices tells you how many attempts a usable clip takes. What generation actually costs explains why that number matters more.Reasoning, September 2026 — compares the FLUX 3 rates above with the rates and clip lengths on this site’s comparison page (Grok Imagine Video 1 to 15 seconds, Gemini Omni Flash 3 to 10 seconds), checked 22 Sep 2026.
Compare like for like: same resolution, same length. Then multiply by the number of tries a usable clip takes.
Are the weights open?
No. The launch plan offers video “through APIs and private weight access”, and lists an open-weight backbone, “FLUX 3 Dev”, as a later step. The model page still says open-weight access “is coming soon.”Black Forest Labs, FLUX 3 - Real World Models, 23 Jul 2026, read at source 22 Sep 2026 · FLUX 3 model page, read at source 22 Sep 2026. Black Forest Labs’ Hugging Face account lists FLUX.1 and FLUX.2 models and no FLUX 3.Hugging Face, black-forest-labs model list, read at source 23 Sep 2026: 35 repositories, none named FLUX 3.
The image side of the family differs: FLUX.2 [klein] 4B has open weights under Apache 2.0, as which image model for which job sets out. That does not apply to FLUX 3.
If you need to run the model on your own machine, FLUX 3 cannot do that yet. Plan around the API, or wait for FLUX 3 Dev.
Can you use the clips commercially?
Yes, according to the developer terms. Black Forest Labs says it claims “no ownership rights in and to your Output”, and that you “may use Outputs for your or their own personal or commercial purposes, subject to any restrictions set forth in these Terms or applicable law.”Black Forest Labs, Developer Terms of Service, last revised 4 Aug 2026, read at source 22 Sep 2026. These are the terms for people outside the EU; the EU version was not read.
Three things come with that permission. The terms say BFL “may embed Content Credentials or other provenance data in any Output”, and you may not remove them. They also warn that using a clip may need third-party rights, “including Audio Rights that we do not hold and cannot grant to you.”Black Forest Labs, Developer Terms of Service, last revised 4 Aug 2026, read at source 22 Sep 2026. And the API terms let the company use your input and output to train: “the Company may use Inputs and Outputs to train and improve its artificial intelligence models”.Black Forest Labs, FLUX API Service Terms, last revised 4 Aug 2026, read at source 22 Sep 2026. For how clauses like these compare across vendors, see who owns what you generate.
You can sell what you make. Do not send footage or audio you are not willing to have used for training, and check the rights to anything recognisable in the soundtrack.
How Black Forest Labs says to prompt it
BFL publishes a video prompting guide separate from its FLUX.2 image guide. The main advice is to lengthen as you go: “Start short to explore an idea, then lengthen the prompt to lock in the details that matter.” For timed action it says: “Keep each beat achievable — two or three for a 5-second clip.”Black Forest Labs, Text-to-Video prompting guide, read at source 22 Sep 2026.
“Put the exact spoken words in quotation marks and say who delivers them.” A visible speaker matters because it “gives FLUX 3 a face to lip-sync”, and for quiet scenes the guide advises: “Name the sound source you want rather than asking for silence.”Black Forest Labs, Audio and speech prompting guide, read at source 22 Sep 2026. The parameter table for the video endpoint has no negative-prompt field. Exclusions such as “No on-screen text” go in the prompt itself, as BFL’s own examples do.Black Forest Labs, Video, parameter table and example prompts, read at source 22 Sep 2026.
Direct the shot rather than describe a picture: what happens, what the camera does, what can be heard.
What this page could not verify
- Quality. Nothing was generated for this page. Black Forest Labs reports that people preferred FLUX 3 to rival models in its own early tests, for example “over Grok Imagine Video in up to 69% of comparisons”, but those are the vendor’s own preliminary tests, not an independent ranking.Black Forest Labs, FLUX 3 - Real World Models, 23 Jul 2026, read at source 22 Sep 2026: “these results are preliminary”.
- An independent ranking. FLUX 3 did not appear on the Artificial Analysis video leaderboards in the copy this site saved on 22 Sep 2026.Artificial Analysis, Text to Video Leaderboard, read in a browser 22 Sep 2026 by this site’s research pass; re-query, rankings change.
- Whether the commit step costs the fhd rate at 2K and 4K. The model page gives the FHD rate for committing a draft, but committing straight to 2K or 4K has been possible since 10 September, and no price is published for that case.
Black Forest Labs: Release notes · API Pricing · FLUX 3 model page · FLUX 3 Video documentation · Video quickstart · Text-to-Video and Audio and speech prompting guides · FLUX 3 - Real World Models (23 Jul 2026) · FLUX 3 Video, Part 1: Generation (4 Aug 2026) · Developer Terms of Service and FLUX API Service Terms (both last revised 4 Aug 2026). All read at source 22 Sep 2026; release notes and pricing re-read unchanged 23 Sep 2026. Hugging Face, black-forest-labs model list, read 23 Sep 2026. Rival prices from this site’s video model price comparison, checked 22 Sep 2026.
Prices and preview terms change. Re-read the pricing page before you budget a job.