FLUX.3 Video API and Pricing, Explained Through a Real Pipeline

FLUX.3 Video is Black Forest Labs' video model, built by the team behind the FLUX image models, and it inherits their priorities. Most video models compete on spectacle. This one competes on control: pin the opening frame, pin the closing frame, and the model works out the journey between them. It also publishes a flat per-second rate at each resolution, so the cost of a clip is the rate times the seconds and nothing else.

This post covers what FLUX.3 Video renders, what each configuration costs, what the endpoint refuses to take, and how to call the model as a published API endpoint instead of building the integration yourself. Every parameter and price below is read off Treza's live video model catalog as of September 2026.

What you get before you talk about price

The interesting part is the combination of length and resolution. Twenty seconds is a long single generation, and holding it at 1080p is rarer still. Seedance 2.5 runs to 30 seconds but tops out at 720p, so for a long take that has to look sharp on a large screen, FLUX.3 Video is the one in the lineup that does both. Six aspect ratios from 21:9 cinema to 9:16 vertical cover a hero film and a phone-first cut from the same model. Full capability detail lives on the FLUX.3 Video model page.

How FLUX.3 Video bills

Per second of output, at a rate set by the resolution and by whether you are generating or continuing:

Those are provider rates you can multiply, which is not something you can say about every video model. ByteDance meters Seedance in opaque video tokens, so working out what a clip costs there means measuring real charges. Here the arithmetic is the whole answer. A 5 second draft at 720p is 85 cents of provider cost, the same 5 seconds at 1080p is $1.45, and a full 20 second take is $3.40 at 720p or $5.80 at 1080p.

Duration is the pricing decision

Because the rate is flat, length is your bill. There is no threshold and no discount at the top of the range: a 20 second generation costs exactly four times a 5 second one.

So draft at 5 seconds. Composition, motion, the travel between your keyframes, and the audio all show up in the first few seconds, and a bad take is bad by second two. Commit to length once the shot is right, and raise the resolution only for takes shown large. The same logic applies one step earlier in the graph, which the first frame post covers: settle the opening image with a cheap image generation before you pay for motion at all.

What you actually pay on Treza

FLUX.3 Video through a Treza pipeline is metered from a prepaid credit balance. Credit packs start at $5, credits never expire, there is no subscription, and a failed generation charges nothing.

One detail is worth knowing before your first run. The platform estimates the media cost of a graph and refuses to start a run your balance cannot cover. FLUX.3 Video has no entry in the estimator's table of verified per-second rates, so it falls back to a conservative floor of 38 cents per second with a $2.26 minimum per clip. A 20 second node holds $7.60 against a published 720p cost of $3.40, and a 5 second node holds the $2.26 minimum against 85 cents.

That hold is not a price. The run settles at actual provider cost plus markup and the difference returns to the balance. The floor errs high on purpose, so a model the estimator does not recognise can never start a job the balance cannot finish.

For sizing a budget across the catalog, the figure on the pricing page is $1.06, the median charge for one video generation across all models. A 5 second FLUX.3 Video draft at 720p lands under it and anything past 10 seconds lands above. The cost breakdown post walks a finished video end to end, with narration, transcription, and captions in the graph rather than just the render.

What the endpoint will not take

No reference images. FLUX.3 Video is not one of the reference-to-video models, so you cannot hand it a set of product or character shots and ask for a clip that matches them. If a character or a product has to look the same across separate generations, use Seedance 2.0 or Seedance 1.5 Pro, both of which accept multi-image references.

That is the counterpart of what the model is good at. Keyframes and references cannot be combined on any model in the catalog anyway, since setting a first or last frame ignores references. FLUX.3 Video picks the keyframe half of that trade.

Audio arrives with the clip, and there is no cheaper silent SKU on the rate card, so a silent draft saves time rather than money.

Continuation, and the honest version of a long take

Continuation is the model's distinguishing mode: hand it footage that already exists and it keeps going, extending the take instead of cutting to a new shot. It has its own line on the rate card at 41 and 53 cents per second, roughly two and a half times the generation rate at 720p, so the provider clearly treats it as the expensive operation.

On a Treza canvas the video node's own inputs are a prompt, a first frame, and a last frame. The shipped path to a piece longer than one generation is assembly: the Sequence node stitches up to 12 shots, and the timeline editor exports up to 40 clips or 15 minutes. Keyframes are what make that assembly hold together, because a shot that has to land on a given composition lands on it rather than near it.

Calling FLUX.3 Video as an API without building the integration

Direct API access to a frontier video model is never one HTTP call. It is credentials, an asynchronous submit-and-poll job pattern, retries, moderation rejections, and somewhere to put the output before your product can use it.

On a Treza AI video pipeline, FLUX.3 Video is a node on a canvas. Set the duration, the aspect ratio, and the keyframes, wire it next to a script step, a voiceover step, or a captioning step, and publish the whole chain as one versioned endpoint. Your application calls a typed /invoke route with JSON in and JSON out, or points an existing OpenAI SDK at the OpenAI-compatible route.

The swap is what pays off later. When the next model wins the leaderboard, you change the dropdown and republish, and the endpoint your product calls does not change. The only thing to check is whether the new model supports the durations and aspect ratios your node was set to, which is where a 20 second FLUX.3 Video node bites you if the replacement stops at 10. Our hosted MCP server exposes the same pipelines as tools, so rendering a brief on FLUX.3 Video and returning the link is something an agent carries out rather than describes.

Where FLUX.3 Video earns its place in a graph

Three jobs. Storyboarded scenes, where somebody has already drawn the opening and closing frame and the generation has to honour both, which is the image to video case. Wide 21:9 footage that has to hold at 1080p, where the alternatives make you choose between length and pixels. And vertical cuts where one 20 second generation replaces a stitch of four, removing three seams you would otherwise have to hide.

The fastest way to answer what it costs for your own work is to run your real prompt at your real duration and read the charge. Open the AI video generator, pick FLUX.3 Video, and generate once at 5 seconds.

Frequently Asked Questions

How much does FLUX.3 Video cost per second?

Seventeen cents per second of output at 720p and twenty nine cents at 1080p, as published for the endpoint we run as of September 2026. Continuation is priced separately at forty one and fifty three cents. Multiply by the seconds for the provider cost of a clip. On Treza that is charged from a prepaid credit balance, and a failed generation charges nothing.

How long can a FLUX.3 Video clip be?

Five to twenty seconds in a single generation, at any whole number of seconds. For longer pieces, stitch generations with the Sequence node, which handles up to 12 shots, or assemble them in the timeline editor, which exports up to 40 clips or 15 minutes.

Does FLUX.3 Video support first and last frame keyframes?

Yes, both ends. Pin the opening frame, the closing frame, or both, and the model generates the motion between them. That is the main reason to pick it over a model that only accepts an opening frame.

Can I use reference images with FLUX.3 Video?

No. It is not a reference-to-video model, so you cannot hand it shots of a character or product and ask for a matching clip. Seedance 2.0 and Seedance 1.5 Pro accept multi-image references if that is the job. No model combines the two: setting a keyframe ignores references.

Does FLUX.3 Video generate audio?

Yes, generated with the picture rather than added afterward. There is no cheaper silent tier on the rate card, so a silent draft costs the same per second.

What is video continuation and what does it cost?

Continuation takes footage that already exists and generates what happens next, preserving the scene instead of cutting away. It is priced above plain generation, at forty one cents per second at 720p and fifty three at 1080p.

Why does my credit balance reserve more than the clip costs?

FLUX.3 Video is not in the estimator's table of verified per-second rates, so it falls back to a floor of 38 cents per second with a $2.26 minimum. A 20 second node holds $7.60 against a published 720p cost of $3.40. The run settles at actual provider cost plus markup and the difference returns to your balance. Reserves are holds, not prices.

Can I call FLUX.3 Video through an API without a Black Forest Labs account?

Yes. On a Treza pipeline, FLUX.3 Video is a node you configure once and publish as a versioned HTTP endpoint, with a typed /invoke route and an OpenAI-compatible route, and a hosted MCP server exposes the same pipelines as tools for agents. There is no subscription: generations are metered from prepaid credits that start at $5 and never expire.