FLUX 3 Max vs FLUX 3: One Page, Two Names, and the Difference You Can Actually Measure

Choosing between FLUX 3 Max and FLUX 3 looks like a choice between two video models. It is closer to a choice between two descriptions of the same model. FLUX 3 Max went live on October 19, 2026 as a named tier of FLUX 3 Video, listed above it on Black Forest Labs' model index, with a speed claim attached and nothing else — no separate price, no separate endpoint, no separate capability list. FLUX 3, the base model that has been generating video since its July 23, 2026 early-access launch, is the thing Max is a faster configuration of. So the useful question is not "which model is better." It is "which parts of the FLUX 3 spec does the Max name change, and which parts does it leave exactly where they were." This piece answers that line by line, because the answer decides whether Max is an upgrade, a premium tier, or a label. 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.

Everything below is read from Black Forest Labs' own model and pricing pages on 2026-10-20, plus the public API schema at api.bfl.ai. BFL publishes no benchmark table for either tier and no Max-specific rate. Where a figure is the vendor's own claim it is labelled that way; where no figure exists, that gap is the finding.

First, what each name refers to

FLUX 3 is the family: one multimodal model that generates video with synchronized audio, generates and edits images, and predicts robot actions, built on an architecture BFL calls Self-Flow. The video tier — FLUX 3 Video — is the part Max is a variant of. FLUX 3 Image, FLUX 3 Image Klein and FLUX 3 Action are the other faces of the same rollout and are not what the Max name touches.

FLUX 3 Max is therefore not a successor and not a new generation. BFL's model index lists it first, with the line "Fastest, highest quality video model. 10-second video in 12 seconds," and links to an anchor on the FLUX 3 Video page rather than to a page of its own. Its product-menu entry points at the same anchor. That structure is the clearest statement of the relationship the vendor makes anywhere: Max is a mode of the video tier, positioned by speed.

The one difference with a number behind it

Speed is the whole of the distinction, and it comes down to four published sentences:

• FLUX 3 Max — "10-second video in 12 seconds," on the model index and on the video page's Max block.

• FLUX 3 Max — "22× faster than Seedance 2.5 at the same quality," repeated in the site announcement bar.

• FLUX 3 Max — "Fastest in the west," which is positioning rather than a measurement.

• FLUX 3 Video — no generation-time figure at all. The base tier's speed story is told through draft mode instead: generate a fast HD preview at a fraction of the cost, then commit the same generation to full quality through Draft Enhance.

That last contrast is the most informative thing on either page. A product that ships a preview-then-commit workflow is telling you that full-quality rendering is the expensive step. A product that claims full-quality ten-second output in twelve seconds is telling you the expensive step is gone. Both statements come from the same vendor, one page apart. They may both be true — Max could be a faster engine sitting behind the same draft UI — but BFL does not say how Max relates to draft mode, and until it does, the two claims should be held separately.

If the twelve seconds holds at your resolution with audio on, the practical effect is large: a ten-second clip takes less wall-clock time than the clip itself, so a shot list becomes something you iterate through rather than queue. If it holds only at HD without audio, the effect is smaller but still real. What BFL has not published — resolution, audio on or off, hardware, queueing, sample size — is exactly the set of details that decides which of those two worlds you are in.

What is identical, and why that is most of the product

The capability envelope is published once, under FLUX 3 Video, and the Max block adds nothing to it. For a buyer this is good news: you are not choosing between feature sets.

• Duration — 5 to 20 seconds per generation, any whole second, or auto. Same ceiling on both, and the API schema enforces the 5–20 integer range on the video endpoint.

• Resolution — HD up to 1 megapixel per frame, FHD up to 2, QHD/2K up to 4, UHD/4K up to 8, bands set by total pixels per frame rather than aspect ratio. Dimensions land on multiples of 32, so 16:9 UHD delivers 3840×2176.

• Inputs — text-to-video, image-to-video, video-to-video, continuation from an existing clip, and keyframe transitions with one start frame, start-and-end pair, or multiple frames pinned to timestamps.

• Audio — native synchronized audio by default, including multilingual dialogue, speech, effects and ambience, at no extra charge.

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

• Multi-shot behaviour — multiple scenes and camera angles inside one generation, which BFL frames as agentic chaining of clips into longer sequences.

• Resolution finishing — FHD and above are the same generation finished by the video upsampler on a larger canvas, so the shot is identical across bands and only the delivered size differs.

If the two tiers share all of that, then Max is a latency choice inside an existing product line, not a different tool. The decision reduces to: do you need the faster render enough to take the less-documented option?

Price: there is no Max rate on the page

The pricing page has one video rate card, and it belongs to FLUX 3 Video:

• 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.

• Worked example — five seconds of standard HD text-to-video is $0.85; a ten-second HD clip is $1.50; partial seconds round up; audio is included.

The page's own footnote says rates per video second are set by modality, by "the model variant selected (e.g., FLUX 3 Video, FLUX 3 Video Draft)" and by resolution band. Max is not among the variants that footnote names, and it has no row. Two readings are possible and BFL does not choose between them: Max costs what FLUX 3 Video costs, or Max carries a premium that has not been published yet. The second is more common in practice for a "highest quality" tier; the first is what the page literally supports. Either way, anyone quoting a Max price today is guessing.

A second complication: BFL's two pages disagree on the base rate. The FLUX 3 Video FAQ on the model page 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. The documentation's billing page quotes $0.15/$0.24/$0.28/$0.41 and $0.21/$0.27/$0.31/$0.45 for the same modes and bands. We lead with the documentation figure because it is the billing reference, and flag the conflict rather than resolve it. Check the invoice.

The API question nobody asking "Max vs FLUX 3" expects

If both tiers were callable, the comparison would end with "pick the faster one." They may not both be callable under separate names.

The public OpenAPI document at api.bfl.ai lists three generation endpoints in 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 accepts mode (t2v, i2v, v2v, draft_enhance), prompt, duration, resolution, aspect_ratio, generate_audio, draft, safety_tolerance, version and user. No field selects Max.

Three explanations fit the evidence, and only BFL can say which is true. Max may be served behind the existing /v1/flux-3-video endpoint with the version parameter or per-account routing deciding the engine. Max may be available only through BFL's dashboard and partner surfaces for now, with API access to follow. Or the documentation has not caught up with a launch made a day earlier — which is the most common explanation for this shape of gap and the least comfortable to build against.

Practically: write your integration against /v1/flux-3-video, keep the model choice in configuration rather than hard-coded, and confirm with BFL which call actually returns Max output before you promise a latency number to anyone downstream.

Where the base tier is still the better documented choice

FLUX 3 Video has three things Max does not, and they are not small:

• A published rate card. You can price a ten-second 4K clip at $4.10 and defend the number. You cannot do that for Max.

• A named endpoint with a documented body. You can write the request today.

• A benchmark history, however thin. BFL's July evaluations compared FLUX 3 against named opponents — Luma Ray 3.2 at 93% preference, Runway Gen-4.5 at 77%, Kling v3 Pro at 60%, Seedance 2.0 and G​emini Omni Flash around 52% — and labelled them preliminary. Max's October claim names no board and no sample. A vendor-run number with a stated method is worth more than a vendor-run rank without one.

None of that makes Max worse. It makes Max newer, and newer means the burden of proof moves to you: run your own prompts, at your own resolution, and measure.

How to decide, in three cases

You are prototyping and iteration speed is the constraint. Try Max first. Even an unverified 12-second render, if it is even close, changes how many directions you can explore in an afternoon. Budget for the possibility that Max bills at the FLUX 3 Video rates and confirm the actual charge on your first invoice.

You are shipping a production pipeline and need a price you can commit to. Build against FLUX 3 Video, whose rate card is published, and treat Max as an experiment behind a feature flag. If Max turns out to be faster at the same rate, promote it; if it carries an unpublished premium, you have lost nothing.

You are choosing a video model against non-FLUX alternatives entirely. The Max-versus-FLUX-3 question is the wrong frame; what matters is that both are one vendor's tiers, and the 22× comparison against Seedance 2.5 crosses two different products with different maximum lengths and reference limits. Compare on your own clips.

Keeping the choice reversible

The awkward part of this decision is that the information needed to make it — Max's price, Max's endpoint, Max's real speed distribution — does not exist yet. That argues for an architecture in which the model choice is a parameter. OrcaRouter provides one OpenAI-compatible endpoint across 200+ models with 0% markup on provider list prices, so when a vendor publishes a new rate or a new tier the number on our side moves with it rather than lagging behind a repricing cycle. Automatic failover matters more than usual when a tier is a day old and unbenchmarked: it lets a request fall through to a proven model when the new one degrades, instead of failing with it. FLUX 3 Max is not routed by OrcaRouter today — for BFL's video API you call the vendor directly — but the models you put around it, including the LLM that writes your shot lists, can sit behind the same key.

The honest close is a short one. FLUX 3 Max and FLUX 3 are the same model at two speeds, or they are the same model and a marketing position — the vendor has not published enough to tell which. What is certain is that the capability envelope did not change, so nothing you learned about FLUX 3 Video's limits stops applying. The tier is a latency bet, and latency bets should be placed with your own stopwatch.