FLUX 3 Video Goes GA: 20-Second Clips, Native Audio

FLUX 3 Video went generally available on August 4, 2026, twelve days after Black Forest Labs first announced the FLUX 3 family. The release ships clips up to 20 seconds, audio generated alongside the frames rather than dubbed on afterwards, and a published price list with four full-render rates plus two draft rates — not the single flat number that will end up in most launch coverage.

The distinction matters more than it sounds. FLUX 3 was announced in late July as a multimodal model spanning image, video, audio and action; what actually opened on August 4 is the video generation leg, via the BFL API and select partners. FLUX 3 Image and an open-weight FLUX 3 Dev variant remain announced roadmap with no dates attached. If you are budgeting or building against this release, you are budgeting against one modality, not the family.

This guide covers what the GA actually includes, the mode-specific limits the launch post glosses over, the full six-tier price structure with our own per-minute conversion so it can be compared against the rest of the field, and an honest read on what the quality claims do and do not establish. Every price is labelled by the surface it came from, because two of them disagree.

Black Forest Labs made FLUX 3 Video generation generally available starting August 4, 2026, through the BFL API and select partners. BFL also states the model went live on Krea the same day; treat that as a vendor statement — no dedicated Krea announcement was locatable at the time of writing, so it is not independently corroborated here.

The headline capabilities are straightforward. Clips run up to 20 seconds. Output is released at HD, meaning 720p natively, with Full HD available through upscaling. Audio is generated alongside the video rather than added in a second pass. And three request modes — text-to-video, image-to-video, and video continuation — share a single flux-3-video endpoint with the same request shape.

On OpenRouter the model is listed as black-forest-labs/flux-3-video, created August 4, 2026, with Black Forest Labs as the only provider — no routing or fallback providers at the time of writing. That single-provider status is worth noting if you route production traffic through an aggregator for redundancy: for this model, right now, there is no second provider to fail over to.

This is a different event from the FLUX 3 announcement in late July, which introduced the model family and framed video as gated early access without pricing. If you want that context, our coverage of the July FLUX 3 announcement covers it; this piece deliberately does not restate it.

The single-endpoint design is genuinely convenient — the same request shape covers all three modes, so switching from a text-driven clip to an image-driven one is a payload change, not an integration change. But the specifications differ per mode in ways the launch post does not spell out, and getting them wrong shows up as failed requests or a blown budget rather than a warning.

The 20-second figure applies to text-to-video and image-to-video. Video continuation is capped at 5 to 15 seconds of output, from up to 4 seconds of supplied footage. Output covers seven aspect ratios — 21:9, 2:1, 16:9, 4:3, 1:1, 3:4 and 9:16 — all at 24 frames per second.

Native audio is the capability BFL leads with, and the framing is specific: sound effects, ambience and dialogue are produced together with the frames, not generated separately and aligned afterwards. The parameter that controls it defaults to on.

Dialogue is lip-synced and multilingual. The launch post names roughly thirteen languages and dialect groups explicitly — English in various dialects, Chinese, Spanish, French, German, Japanese, Portuguese, Russian, Italian, Indonesian, Turkish, Hindi and Punjabi — followed by “and more.” Read that as about thirteen named with an open tail, not as a certified count; BFL has not published a definitive language list.

The second capability worth separating out is multi-shot generation. FLUX 3 Video can produce multiple scenes and camera angles inside a single generation while keeping the sequence coherent. That is a single-request capability, not an editing workflow — which is a meaningful difference for anyone who has built a pipeline that generates shots individually and then fights continuity drift in the edit.

Here is the trade-off nobody puts in a launch post. Joint audio and video generation removes an entire production step, and it also removes an entire control surface. When the soundtrack is baked in at generation time, you cannot swap the voice, re-time the music, or replace a sound effect without regenerating the clip — and regenerating means paying again and accepting new frames. For social-first content and explainer video that is a clean win. For work where audio is licensed, legally reviewed, or brand-locked, the practical pattern is to generate with audio off and keep the sound design in a pipeline you control.

BFL also positions the model for documentary and short-form educational content generated from short prompts, citing pretraining world knowledge combined with real-time grounding. That is vendor positioning rather than a benchmarked capability, and it is the claim most worth testing on your own prompts before you build a content line on it.

The number that will travel is $0.17 per second. It is real, and it is the cheapest full-quality tier: text-to-video or image-to-video at HD. It is also one of six published rates. Full HD full renders go to $0.29 per second. Video continuation is priced on its own ladder at $0.43 per second HD and $0.54 per second Full HD on BFL list. Draft Mode sits underneath everything at $0.06 per second for text-to-video and image-to-video, and $0.12 per second for continuation.

Run the spread and the shape of the pricing becomes obvious. From the cheapest published rate to the dearest is a factor of nine. Even restricting to full renders, continuation at Full HD costs 3.2× what an HD text-to-video second costs. A pipeline that treats “FLUX 3 Video” as one line item in a spreadsheet will misprice itself by multiples depending on which modes it actually leans on.

Draft Mode deserves its own paragraph because it is the lever that changes unit economics rather than shaving them. A draft returns a fast, low-cost preview of the prompt; once a draft is approved, draft-enhance renders the same subjects, composition and motion at full quality from the draft cache bundle rather than reinterpreting the prompt from scratch. BFL describes drafts as costing about a third of a full render, which is almost exactly right against the HD tier and understates the saving at Full HD.

Put the draft mechanic into a realistic iteration loop and the case makes itself. Eight full-length draft passes plus one HD full render at 20 seconds costs $13.00 — eight at $1.20 plus $3.40. Nine HD full renders, which is what the same amount of exploration costs without drafts, comes to $30.60. That is $17.60 saved on a single shot, or 57% off the exploration budget, and the saving scales linearly with how many variants your process actually needs before someone signs off.

Comparing video models on price is harder than it should be because the industry quotes in two incompatible units. BFL prices per second by mode and resolution. Independent leaderboards price per minute of finished video — Artificial Analysis states its API pricing column reflects the cost to generate one minute of 1080p video on the model creator’s API at default settings. Nothing lines up until one side is converted.

The table below does that conversion. Every FLUX 3 Video row starts from a published per-second rate and multiplies by 60; every competitor row starts from a listed per-minute price and divides by 60. The direction of each conversion is marked, because a derived number should never be mistaken for a quoted one.

The like-for-like row against Artificial Analysis pricing is FLUX 3 Video at Full HD, $17.40 per finished minute, because that is the only tier that matches the 1080p basis the leaderboard uses. On that axis FLUX 3 Video comes in 27.5% below Veo 3.1 at $24.00 and 13.7% below Kling 3.0 Pro at $20.16 — but 2.2× above MiniMax H3 at $7.80 and 2.9× above Gemini Omni Flash at $6.00, which happens to be the model sitting at the top of the text-to-video Arena.

Read that honestly and the positioning is mid-market, not price-leading. FLUX 3 Video undercuts the incumbent premium tier and sits well above the aggressive Chinese and Google-efficiency tiers. The exception is Draft Mode: at $3.60 per finished minute it is 25% below the cheapest scored row in the table, Veo 3.1 Lite at $4.80. That puts BFL’s cheapest published rate on iteration rather than on final renders.

For the wider economics of per-second video pricing and where these rates sit against the cost of the ad creative they replace, we unpacked the trend line in the per-second video pricing war, and the Kling side of the table is covered in our Kling 3 guide.

Every video launch arrives with a preference chart, and this one is no exception. What matters is reading the vendor’s own wording rather than the headline someone writes on top of it.

Three things follow from that sentence. First, it is an internal evaluation, so the rater pool, prompt set and comparison conditions are the vendor’s own. Second, on image-to-video the claim is a tie with Seedance 2.0 and a win over everyone else — not a win over Seedance. A great deal of launch coverage will round that up, and it is wrong to. Third, the comparison names Seedance 2.0, not Seedance 2.5, which is a different and newer model.

BFL published a rater-preference chart alongside the claim, but its headline figure appears in the chart caption rather than the body text, and it is a score from the vendor’s own evaluation rather than a public arena. We are not printing that number as a benchmark, for the same reason we would not print an internally-scored exam grade as an external qualification.

The independent picture is simply absent. At the time of writing, FLUX 3 Video had no row on either the text-to-video or image-to-video Artificial Analysis Arena leaderboard. Neither did Seedance 2.5. Both are too recent for a voting-based arena to have accumulated comparisons, which is a limitation of the method rather than a judgement on either model. For reference, the top of the text-to-video-with-audio board at the time of writing was Gemini Omni Flash at 1244 Elo, with MiniMax H3 second at 1238 and Seedance 2.0 720p third at 1224 — that is the band a new entrant needs to reach to be competitive on measured preference rather than vendor preference.

One piece of pre-release diligence is worth noting because it is rarely disclosed at all: BFL states it worked with third-party partner Cinder to evaluate FLUX 3 Video across supported modalities before release, including for non-consensual intimate imagery and CSAM risk. That is a process disclosure rather than a published result, but for agencies with brand-safety obligations, a named external evaluator is more than most video launches offer.

Our pre-GA snapshot of this field — written before H3 shipped and before Seedance 2.5 launched — is at the earlier video field comparison. This post supersedes its pricing and availability rows.

Price is one axis; what the model will let you build is another. The matrix below covers the three models with a primary specification source retrieved for this piece — FLUX 3 Video, MiniMax H3 and Seedance 2.5. Cells that the primaries do not establish are marked as such rather than filled in from secondary coverage.

Two availability notes change how these rows should be read. MiniMax H3 launched at the end of July with a live pay-as-you-go API, and we covered it in the MiniMax H3 launch write-up. Seedance 2.5 launched into consumer surfaces with API access described as coming soon in its own announcement, covered in the Seedance 2.5 one-take launch. A 30-second one-take capability you cannot yet call from an API is not, today, a pipeline option — which is the practical reason FLUX 3 Video and H3 are the two rows that matter for anyone building this quarter.

The routing question is not “is FLUX 3 Video the best model” — with no independent score, that question has no defensible answer yet. The useful question is which jobs its specific shape fits, and where something cheaper or longer-running fits better.

The workflow that falls out of the pricing is unusually explicit for a video model: iterate in Draft Mode, approve a shot, then use draft-enhance so the full render reproduces the approved subjects, composition and motion from the draft cache rather than reinterpreting the prompt. That is the difference between paying for exploration once and paying for it every time the model reinterprets a brief. Teams that skip drafts because the per-second rate looks cheap will spend multiples of what a draft-first process costs.

Where this heads next is worth stating as a projection rather than a fact. The pattern across this year’s video launches is that the headline specification — clip length, resolution, native audio — converges quickly, while the pricing structure is where vendors differentiate. BFL has priced a deep preview tier into this release, and the plausible next move is for iteration cost to become a pricing battleground in its own right, because it is what actually determines whether AI video is viable at catalogue or campaign scale. Independent Arena scores for both FLUX 3 Video and Seedance 2.5 should appear within weeks; until they do, treat every quality ranking in circulation, including the vendor’s, as provisional.

If you are pricing this into a campaign rather than a prototype, the catalogue-scale economics get their own treatment in our guide to product video at catalogue scale, and the media-side question of what generated video actually does to creative performance sits inside our paid media engagements. For teams building the generation pipeline itself rather than buying the output, our ecommerce practice runs the SKU-consistency and cost-modelling side.