FLUX 3 vs Seedance, Omni & Kling: Mapping the Video Field

The best AI video model in July 2026 depends entirely on which axis you measure. FLUX 3 landed on July 23 with the longest single-pass clip Black Forest Labs has ever published — 20 seconds with native audio — but no pricing, no independent benchmark, and gated access. Seedance 2.5 and Gemini Omni Flash, meanwhile, are shipping in production today.

The stakes are practical, not academic. Marketing and creative teams are budgeting real money against per-minute API rates that range from $6.00 to $24.00 on the Artificial Analysis leaderboard, while two of the most-hyped models in the field — FLUX 3 and Seedance 2.5 — publish no official per-minute price at all. Vendor comparison charts all favor their own model, and almost none put duration, price, control surface, and legal risk on the same page.

This guide does exactly that. We map six models — FLUX 3 Video, Seedance 2.5, Gemini Omni Flash, HappyHorse 1.1, Kling 3.0, and Veo 3.1 — across every axis a buying team actually weighs, mark every unverifiable cell as “not published” rather than guessing, and close with a which-model-for-which-job decision matrix.

Two launch events frame this comparison. On July 16, ByteDance opened public API access to Seedance 2.5 via BytePlus — completing the rollout announced at Volcano Engine’s FORCE conference on June 23. Seven days later, on July 23, Black Forest Labs announced FLUX 3, its first multimodal frontier model, jointly trained across image, video, and audio in one architecture. Between those two dates sits a field that already looked crowded: the post-Sora AI video market has been consolidating fast since OpenAI discontinued the Sora consumer app (announced March 24, offline April 26), with the Sora API itself sunsetting September 24, 2026.

That leaves six models worth a marketing team’s attention today — and three distinct storylines among them.

FLUX 3 is Black Forest Labs’ first multimodal frontier model — jointly trained across image, video, and audio rather than bolting a video head onto an image backbone. The headline capability: up to 20 seconds of video with native audio in a single generation, the longest single-pass duration BFL has published. Preliminary evaluations were run at 720p on 10-second clips. For the full architecture story, launch tiers, and the robotics angle, see our full FLUX 3 launch breakdown.

What shipped is narrower than the announcement suggests. FLUX 3 Video is in gated Early Access — API plus private weights by request. FLUX 3 Image is “coming in weeks” and had not shipped as of July 24. FLUX 3 Action is limited to selected robotics partners, and FLUX 3 Dev — the open-weight multimodal backbone — is a future release with no date. Martin Scorsese reportedly joined BFL as an advisor in early July, a signal of how seriously the lab is courting film-adjacent credibility. Early testers named at launch include Canva, Burda, Magnific, Krea, and Picsart.

The only comparative data available is BFL’s own preference testing — and BFL itself labels the chart a “preliminary evaluation of an early FLUX 3 candidate,” meaning the tested checkpoint is not necessarily the model now in Early Access. Read the numbers with that caveat attached: the blowout wins come against the field’s trailing models, while against the two models that actually lead the arena — Seedance 2.0 and Gemini Omni Flash — FLUX 3 scores 52%, a statistical coin flip.

Seedance 2.5 generates a native, unstitched 30-second clip in a single API call — the longest single-pass commercial video generation available as of the July 16 API opening. For context from TechTimes’ read of the competitive field: Veo 3.1 caps native passes at 8 seconds and Runway Gen-4.5 at 5–10 seconds. We covered Seedance 2.5’s 30-second single-pass launch in depth when the rollout was announced; the API opening makes it real for production teams.

The control surface is the other differentiator. Seedance 2.5 accepts up to 50 multimodal reference inputs — images, video clips, and audio — in one generation request, up from 15 total in Seedance 2.0 (9 images, 3 video, 3 audio). That is more than a threefold jump, and it matters most for brand work: product shots, brand footage, and voice references can all steer one generation. Region-level editing is new too — modifying a specific face, object, or background element in an already-generated clip without regenerating the whole shot.

Under the hood, ByteDance describes a Sparse Diffusion Transformer built by its Doubao team, using sparse attention across the full clip duration rather than a sliding window to hold character, lighting, and camera-position consistency over 30 seconds, with audio co-generated in the same latent space as the video rather than post-synced. Treat those as vendor-stated specs: no independent third-party benchmark of Seedance 2.5 existed as of the API opening, and the model has no confirmed Artificial Analysis leaderboard entry — the arena still lists Seedance 2.0.

One number Seedance 2.5 does not have: a price. ByteDance has published no official per-minute API rate for Seedance 2.5 as of this writing. The Artificial Analysis leaderboard entry at $9.07 per minute is for Seedance 2.0 720p — a different model — and third-party reseller estimates for 2.5 are not ByteDance’s published rate. In our comparison table below, that cell reads “not published,” because it is.

If Seedance owns duration, Gemini Omni Flash owns everything measurable. As of July 24, it is ranked #1 on all four Artificial Analysis Video Arena boards: Text-to-Video with audio (Elo 1,245), Text-to-Video without audio (1,326), Image-to-Video with audio (1,201), and Image-to-Video without audio (1,375). It is also the cheapest top-tier model on the leaderboard at $6.00 per minute for 1080p API generation — $0.10 per second, or roughly $1.00 for a 10-second 720p clip per VentureBeat’s pricing table. Since Gemini Omni’s video-generation debut, Google has merged generation and conversational editing into one interface, which is where the workflow advantage compounds.

The constraints are equally concrete. Clips run 10 seconds maximum (3-second minimum) at 720p/24fps — a 30-second social cut means three generations and a stitch. And editing of uploaded (rather than model-generated) video is unavailable in the EEA, Switzerland, and the UK — a real limitation for European teams hoping to iterate on existing brand footage.

Read the price chart alongside the ranks and the field inverts the usual assumption that quality costs more: the #1-ranked model is the cheapest, and the most expensive model in the set ranks eleventh. Veo 3.1’s premium buys native 4K and per-second billing, not arena preference. For high-volume iterative work — concept testing, storyboard animatics, ad-variant drafts — Omni Flash at $6.00 per minute is four times cheaper than Veo 3.1 ($24.00 ÷ $6.00 = 4×) and undercuts every AA-listed rival in this set.

Alibaba HappyHorse 1.1 (June 23, 2026) ranks #4 on the AA Text-to-Video-with-audio board (Elo 1,149) and #4 on Image-to-Video with audio (1,110), at $9.90 per minute. It generates 15-second clips at 1080p/24fps with native audio — lip-sync plus Foley. Its predecessor, HappyHorse 1.0 (April 2026), still ranks #5 (Elo 1,128) at $13.20 per minute, and reportedly took the #1 arena spot within days of its April 7 debut, beating both Seedance 2.0 and Kling 3.0 at the time. Alibaba’s Wan2.7 line holds #3 (Elo 1,164) on the same board — we unpack Alibaba’s two-line video stack in a companion post. The oft-repeated claim that HappyHorse was built under Zhang Di, the engineer behind Kling 1.0 and 2.0, comes from aggregator coverage rather than Alibaba itself — treat it as reported, not confirmed.

Kling 3.0 (Kuaishou, released February 2026 — not a new launch this window) sits mid-pack: the Pro 1080p tier ranks #6 (Elo 1,111, $20.16/min), Standard 720p #9 (1,100, $15.12/min), with the Omni variants at #10 and #12. That makes Kling the most expensive per-minute line in this comparison except Veo 3.1. Its control-surface gap is stark: per TechTimes’ competitive read, Kling 3.0 accepts text and image inputs only — no video or audio reference input — against Seedance 2.5’s 50-slot multimodal system.

Google Veo 3.1 remains the production Veo line — no new Veo number shipped in this window. It ranks #11 (Elo 1,095) at $24.00 per minute, the most expensive model in the set, but offers native 4K and per-second billing, which is why it keeps its place in premium broadcast workflows. Veo 3.1 Lite (#13, Elo 1,092, $4.80/min) and Veo 3.1 Fast (#14, Elo 1,091) fill out the budget tiers — with a pricing wrinkle on Fast we address under the table below.

No published source puts all six models on one table with duration, audio, control, arena rank, price, and availability together — vendor pages show only their own model, and the AA leaderboard has rank and price but not duration or control detail. This table is the union. Every cell traces to a cited source; where no official figure exists, the cell says so rather than guessing.

One cell we deliberately left out of the table: Veo 3.1 Fast. Its two sourced prices do not reconcile. Artificial Analysis lists $9.00 per minute; VentureBeat’s per-clip table implies roughly $6.00 per minute ($1.00 per 10-second clip × 6). For Veo 3.1 itself the same cross-check reconciles cleanly ($4.00 per 10 seconds × 6 = $24.00/min, matching AA), so the Fast discrepancy is real, not a rounding artifact. We present both figures with their sources rather than picking one.

The trend the table exposes: the field is racing on different axes because nobody can win them all at once. Long single passes demand attention architectures that hold consistency over 30 seconds; cheap iteration demands short clips and aggressive serving economics; premium broadcast demands 4K pipelines. Each vendor has optimized for the axis its distribution favors — ByteDance for social-native long takes, Google for volume and integration, Kuaishou and Alibaba for the Chinese creator stack. Expect the axes to converge slowly, not suddenly: the physics of serving cost means the 30-second-single-pass and $6-per-minute clubs will stay separate clubs for a while.

Most “which video model” comparisons omit legal exposure entirely. For an agency advising clients, it belongs in the table. Seedance leads on spec but carries the field’s most concrete legal risk: after Seedance 2.0’s February 12, 2026 China launch went viral with unauthorized celebrity likenesses, five major studios — Disney, Warner Bros. Discovery, Paramount Skydance, Netflix, and Sony Pictures — sent cease-and-desist letters, and on February 22 the Motion Picture Association sent its first-ever AI cease-and-desist to a major generative-AI company.

As of July 2026, no federal lawsuit has been filed — serving a Beijing-headquartered company through the Hague Service Convention is estimated at 18–24 months — which means the exposure is unresolved rather than resolved. Industry sentiment captures the ambivalence: Seedance has been described by media advisor Peter Csathy as “the most powerful video generator in the market right now,” even as studio approval remains informal at best.

Vendor comparisons all favor their own model. A marketing-team lens routes by workload instead — here is how we would assign the field today, with the caveat that every recommendation assumes you run a small paid pilot on your own briefs before committing a quarter’s production budget.

And FLUX 3? Until pricing, open access, and an independent benchmark exist, it is a watch-list item, not a production decision. The 20-second single-pass claim with native audio is genuinely differentiated — but a 52% coin flip against the two arena leaders, from a vendor-run preliminary eval of a checkpoint that may not match the shipping model, is not a basis for moving budget. If you want help wiring any of these models into a real pipeline — creative testing on paid channels, always-on social production, or cost modeling per deliverable — our social media team runs exactly these evaluations, and our paid media practice can pressure-test whether AI video actually lifts your creative performance before you scale it.