FLUX 3 Max: 10 Seconds of Video in 12 Seconds, and a Claim You Cannot Check Yet

The most interesting number in Black Forest Labs' FLUX 3 Max announcement is not 22×. It is 12. A ten-second clip in twelve seconds means the model spends less time rendering than the clip spends existing — and if that holds in production, the entire shape of a video workflow changes, because iteration stops being a batch job and becomes a conversation. Here is the part worth reading before you rebuild anything around it: FLUX 3 Max shipped on October 19, 2026, the 12-second figure is the vendor's own, published with no methodology, no clip size, no resolution and no independent run behind it, and the model has no price row of its own on Black Forest Labs' pricing page. What FLUX 3 Max is may be the most important thing to understand, and it is the thing the announcement page says least about. This piece covers what BFL actually shipped, what the numbers mean and do not mean, what it costs, and how it relates to the FLUX 3 Video tier you could already call. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

A note on sourcing, before any number appears twice. Everything here is read off Black Forest Labs' own pages on 2026-10-20 — the model page at bfl.ai/models, the FLUX 3 Video product page, and the pricing page — plus the OpenAPI schema of the public API. No third party has benchmarked FLUX 3 Max. Where BFL publishes a number, it is labelled as the vendor's. Where it publishes nothing, this article says so instead of estimating.

What was actually announced

On October 19, 2026, Black Forest Labs posted a launch note for FLUX 3 Max under the line "Introducing FLUX 3 Max." The claims in it are short and total: fastest in the west; a ten-second video in twelve seconds; tied for first in image-to-video; 22× faster than Seedance 2.5 at the same quality; and, in a sentence that reads like it belongs to a different product, "FLUX 3 is a model designed for customization."

What the company simultaneously did elsewhere is more useful than the post. FLUX 3 Max now occupies the first slot on bfl.ai/models, above FLUX 3 Video, with the description "Fastest, highest quality video model. 10-second video in 12 seconds." It has its own product-menu entry and its own anchor on the video page. The pricing page carries a site-wide announcement bar for it: "FLUX 3 Max: 22× faster than Seedance 2.5 at the same quality."

That is a real product launch in the sense that matters to a developer: the model is listed, it has a name, and the vendor is steering traffic to it. It is not a launch in the sense of a new architecture, a new modality or a new price tier. FLUX 3 Max is a variant of FLUX 3 Video — and the distinction between "new model" and "new variant" is where most of the practical questions live.

The 12-second claim, and what it is measured against

Twelve seconds for ten seconds of output is a ratio of 1.2 to 1. If it is true at production resolution with audio, it is extraordinary. For comparison, the vendor's own FLUX 3 Video family page describes draft mode as existing precisely because full-quality renders are not instant — a draft buys you "fast generation at a fraction of the cost," and you then commit the draft back through Draft Enhance to get the full-quality frame. A workflow built around preview-then-commit implies that the commit is the slow part. FLUX 3 Max's claim implies the opposite: that the full render is already faster than real time.

Four things BFL does not say about the twelve seconds, and each one changes what the number means:

• Resolution — the claim does not name a resolution band. FLUX 3 Video renders HD, FHD, QHD and UHD, with FHD and above finished by an upsampler after generation. A twelve-second HD render and a twelve-second 4K render are different engineering achievements, and the post does not distinguish them.

• Audio — FLUX 3 generates synchronized audio by default, and audio is included at no extra charge. Whether the twelve seconds includes audio synthesis is unstated. Audio is a meaningful share of the compute in a joint audio-video model.

• Hardware and queue — no GPU, no batch size and no mention of whether queueing is included. Time-to-first-frame and time-to-finished-file are different numbers, and marketing speed claims usually quote the better one.

• Repetition — no p50, no p95, no sample size. One fast generation is an anecdote; a distribution is a service level.

None of this makes the number false. It makes it unverifiable from outside the vendor, which is the honest status of every speed claim published in the first 48 hours of a model's life.

"Tied for #1 in image-to-video" — tied with whom, on what

This is the claim to treat with the most care, because it is the one that sounds like a benchmark result and is the one with the least behind it.

A rank requires a board. As of 2026-10-20, the independent video leaderboards do not obviously carry FLUX 3 Max: Artificial Analysis's video arena, which does publish an image-to-video tab with Elo-style rankings, presents the voting interface and its model pool to a signed-in visitor, and the public page reachable without an account does not display the ranked table. So the claim cannot be checked against the most likely source, and BFL does not name the source it is using.

There is also a date problem. A model launched on October 19 cannot have accumulated a meaningful blind-vote sample by October 20. If the "#1" comes from a vendor-run human-preference comparison — which is exactly how BFL presented FLUX 3 Video in July, when it reported preference wins over Luma Ray 3.2 at 93%, Runway Gen-4.5 at 77% and Kling v3 Pro at 60%, explicitly labelled "preliminary" and "early" — then "tied for #1" is a vendor-run result wearing a leaderboard's clothes. The July numbers were honest about their own provisional status. The October post is not, and that is the difference worth flagging.

What would settle it: a named board, a date, a sample size, and a link. Until one appears, the fair reading is that FLUX 3 Max is positioned at the top of the image-to-video field by its maker and has not been independently placed there.

The 22× Seedance comparison

22× faster than Seedance 2.5 "at the same quality" is a compound claim with two halves, and only one of them is a speed claim.

The speed half is arithmetic on the twelve-second figure: if FLUX 3 Max renders ten seconds in twelve, then matching output from a model 22× slower takes about four and a half minutes. That is a plausible order of magnitude for a frontier video model, and it is also completely unanchored — ByteDance publishes no generation-time figure for Seedance 2.5 that would let anyone check it. BFL's comparison is therefore a vendor measurement of its own latency against an unnamed measurement of a competitor's, and the reader has no way to reproduce either side.

The quality half — "at the same quality" — is the half with no evidence attached at all. Matching quality is the expensive part of a speed comparison, and asserting it without a matched harness makes the 22× a single number pretending to be two. Note also that Seedance 2.5 is a different kind of product: ByteDance's model accepts up to 50 reference assets in one request and generates clips up to 30 seconds, against FLUX 3's 20-second ceiling. A speed ratio between models with different length and reference limits is a ratio between two different jobs.

This does not mean FLUX 3 Max is not fast. It means 22× is a vendor comparison between a vendor's own measurement and a competitor's published settings, and the reader should hold it as a direction, not a multiplier.

What FLUX 3 Max actually generates

Here the picture is clearer, because FLUX 3 Max inherits the FLUX 3 Video envelope rather than defining a new one. BFL's video page lists the capabilities under FLUX 3 Video, and the Max block on that page adds only speed language on top of them:

• Inputs — text-to-video, image-to-video, video-to-video, video continuation, and keyframe-controlled transitions.

• Duration — 5 to 20 seconds per generation, any whole second, or auto to fit the content. The API schema confirms the 5–20 integer range.

• Resolution — HD (up to 1 megapixel per frame), FHD (up to 2), QHD/2K (up to 4), UHD/4K (up to 8), where the band is set by total pixels per frame rather than aspect ratio. Delivered dimensions land on multiples of 32, so 16:9 "UHD" is 3840×2176 rather than 3840×2160.

• Audio — optional native synchronized audio, including multilingual dialogue, speech, sound effects and ambience, at no extra charge. On by default in the API.

• Aspect ratios — auto, 21:9, 2:1, 16:9, 4:3, 1:1, 3:4, 9:16, 9:21.

• Reference and control modes — start frame, start-and-end frame, multiple keyframes pinned to timestamps, and continuation from an existing clip.

• Draft mode — HD-only fast preview, with a draft_cache bundle you pass back to render the same generation at full quality. BFL's own framing of draft mode is the tell: if Max rendered everything in twelve seconds, a preview tier would have little to do.

The API surface is the part worth checking before anyone plans an integration. The public OpenAPI document at api.bfl.ai lists three generation endpoints for this family: /v1/flux-3-video, /v1/flux-3-image and /v1/flux-3-image-klein. There is no /v1/flux-3-max. The video endpoint's request body carries mode, prompt, duration, resolution, generate_audio, draft, safety_tolerance and version — and no field that selects Max. FLUX 3 Max is, as of this reading, a product and a placement rather than a distinct, publicly documented endpoint. Anyone building against it should confirm with BFL which endpoint and which parameter actually serve it before writing code that assumes a model ID.

What it costs

Black Forest Labs publishes per-second video pricing, and FLUX 3 Max is not on the rate card. The billing reference — the documentation's pricing page — lists the FLUX 3 Video rows and nothing under the Max name:

• Text-to-video and image-to-video — $0.15/s HD, $0.24/s FHD, $0.28/s QHD, $0.41/s UHD.

• Video-to-video and continuation — $0.21/s HD, $0.27/s FHD, $0.31/s QHD, $0.45/s UHD.

• Draft — $0.06/s, HD only, text-to-video only; a draft is then enhanced at the rate for the band you commit to.

• Worked example — five seconds of standard HD text-to-video is $0.85; partial seconds round up; audio is free.

Two things follow. First, a ten-second HD clip at the listed rate is $1.50, and if Max is priced at those same rates then the twelve-second render costs a dollar fifty plus twelve seconds of your time. Second, BFL's own pages do not agree with each other: the FLUX 3 Video FAQ on the model page still quotes $0.17/s HD, $0.29/s FHD, $0.40/s QHD and $0.80/s UHD for text-or-image-to-video, with $0.41/$0.53/$0.65/$0.95 for video-to-video — roughly double the documentation's card at every band. We quote the documentation page because that is the billing reference, and we flag the conflict rather than pick a winner. Anyone budgeting should confirm the rate on a first invoice before scaling. The same footnote also says rates are determined by modality, model variant and resolution band — which implies a Max-specific rate exists somewhere, even though no Max row is published. Treat $0.15/s HD as an assumption, not a Max quote.

For a team comparing options across vendors, the useful property of a pass-through router is that this kind of rate-card churn is not your problem to track: on OrcaRouter, provider list prices are passed through with 0% markup, so a vendor cut or a new tier's rate appears at the same price it appears on the vendor's page, on the same day, without a repricing exercise on your side. FLUX 3 Max itself is not routed by OrcaRouter today — BFL's video API is the route — but the neighbouring video options that are on the catalogue, such as kling/kling-3-turbo, sit behind the same OpenAI-compatible endpoint as the text models, so a pipeline can keep its LLM and video calls on one key while it evaluates.

Customization, and the open-weights thread

The customisation sentence in the launch post is the one claim that does not fit the rest of the announcement, and it is worth pulling apart because BFL has actual form here.

The FLUX 3 family has been opening up in pieces all year. FLUX 3 Action, the 7B robot-action model, published three checkpoints to Hugging Face on September 22, 2026. FLUX 3 Image Klein, a 5.6B open-weights image model, shipped on October 15 with a base checkpoint built for fine-tuning, a distilled variant, FP8/INT8/NVFP4 builds, LoRA training in Diffusers and Ostris's AI Toolkit from day one, and a FLUX Kommunity License that permits free commercial use below $5M annual revenue. BFL's own blog calls Klein "the second open weights release from the FLUX 3 family" and says more will follow.

Against that record, "FLUX 3 is a model designed for customization" is plausibly a statement about the family — the image side is literally fine-tunable, on your own GPU, today — rather than a claim that FLUX 3 Max weights are coming. Nothing on the Max page says weights, licence, fine-tuning or self-hosting. The FLUX 3 Dev video backbone is still described as "coming soon," and bfl.ai/licensing sells self-host tiers only for the image models. If customisation is why FLUX 3 Max interests you, the honest position is that the video tier's customisation story is a vendor sentence with no artefact behind it yet, while the image tier's is four checkpoints and a licence you can read.

What to do with it this week

If you generate video for a living, FLUX 3 Max is worth a trial run on your own prompts — not because the 22× has been verified, but because a genuine sub-real-time render changes how you work, and the only way to find out whether the claim holds for your resolution, your audio settings and your content is to send it your own material. Benchmark it the way the vendor did not: same prompt, same duration, three resolutions, ten runs each, and record p50 rather than the best case.

If you are choosing a video vendor for a product, the information you need does not exist yet. There is no independent image-to-video placement, no Max-specific rate, and no documented Max endpoint. Ask BFL directly which endpoint parameter serves Max and what it costs at each resolution band before you size a budget around it. A model that leads a leaderboard is not the same as a model with a published price, and right now FLUX 3 Max has the first claim and not the second.

If you are building a multi-model pipeline, the practical move is to keep the model choice behind an interface rather than in the code. That is the argument for a router that is not specific to this launch: video model generations are turning over every few weeks, and the cost of adopting the next one should be a configuration change. On OrcaRouter that means a single OpenAI-compatible endpoint across 200+ models, automatic failover when a provider degrades, and a routing DSL that can send a class of requests to one model and the rest to another — so evaluating FLUX 3 Max alongside what you run today does not require a second integration or a second contract.

The claim to watch is not the 22×. It is whether an independent board lists FLUX 3 Max in image-to-video, and whether BFL publishes a Max row on its pricing page and a Max parameter in its API. Those three artefacts would turn the launch post into a specification. Until then, the twelve seconds is a promise, and the most useful thing you can do with a promise is test it.