R4. AI video models and pricing, as of 2026-09-04

Source of truth: fal's model API queried with the project key on 2026-09-04 (154 text-to-video, 233 image-to-video, 223 video-to-video endpoints), plus fal model pages, lab blogs, and X posts fetched the same day by a research agent. Prices are fal page text as of today. Promo prices expire: H3 Max Sep 7, H3 Max Director Sep 14.

Flagship models

Lab / model	Release	fal endpoint	fal price per second	Max clip	Max res	Native audio	Ref-to-video	Multi-shot	Source
Alibaba Wan 3.0 Prime	GA Aug 24 2026	`alibaba/wan-3.0-prime/text-to-video`, `/image-to-video`, `/reference-to-video`; also `alibaba/wan-3.0/*`	$0.068 480p, $0.14 720p, $0.28 1080p	2 to 30s single pass	1080p	Yes, same pass	10 img + 5 video + 5 audio + doc/webpage	Long take	Reuters, fal
ByteDance Seedance 2.5	Jul 31 2026; fal Aug 7	`bytedance/seedance-2.5/text-to-video`, `/image-to-video`, `/reference-to-video`	~$0.22 480p, ~$0.47 720p, ~$1.16 1080p (token billed, $0.0214 per 1k)	4 to 30s single pass	1080p (720p recommended)	Yes, joint latent	Up to 50 refs (30 img, 10 video, 10 audio)	Yes, single pass	Seed blog, fal
MiniMax H3 Max (fal post-trained)	Aug 26 2026; PR Sep 1	`minimax/h3-max/text-to-video`, `/image-to-video`, `/reference-to-video`, `/director`; `minimax/h3-max-turbo/*`	$0.0125 480p, $0.02 768p promo to Sep 7 (list $0.05 / $0.08); Turbo $0.01 768p; Director $0.02 promo, 60s minimum	5 to 15s	768p on fal	Yes, stereo	Yes	Director endpoint	MiniMax docs, fal, PR
MiniMax H3 (open weights)	Jul 30 2026	`minimax/h3/text-to-video`, `/image-to-video`, `/reference-to-video`, `/text-to-video/lora`, `/image-to-video/lora`	higher than H3 Max per X reports ($1.30 vs $0.20 per clip)	4 to 15s	2K	Yes	Yes	No	kfgo
Google Gemini Omni 1.1 Flash	Omni Flash May 19 2026; 1.1 Aug 27 2026	`google/gemini-omni-flash/v1.1/text-to-video`, `/image-to-video`, `/reference-to-video`, edit	$0.03 360p, $0.10 720p, $0.15 1080p, $0.30 4K	8s default; extend to 40s per Google	4K	Yes	Yes	Via extend	Google blog, fal
Google Veo 3.1 (Lite tier Apr 2026; no Veo 4 found)	Oct 2025	`fal-ai/veo3.1`, `/fast`, `/lite`, `/reference-to-video`, `/image-to-video`, `/first-last-frame-to-video`, extend	$0.20 silent, $0.40 audio (720/1080); 4K $0.40 / $0.60; fast $0.10 / $0.15	8s; extend to ~148s	4K	Yes	Yes	Via extend	fal, Vertex Lite
Kuaishou Kling 3.0 / 3.0 Omni (O3); 3.0 Turbo Jun 17; native 4K Apr 23	Feb 5 2026	`fal-ai/kling-video/v3/{standard,pro}/`, `o3/`, `v3/turbo/`, `v3/4k/`	$0.112 silent, $0.168 audio, $0.196 with voice control	3 to 15s	1080p (4K request types exist)	Yes (EN/ZH voice)	Elements on i2v / o3	`multi_prompt` 1 to 6 shots within 15s	Kuaishou PR, fal
Lightricks LTX 2.5 (open weights)	Aug 11 2026	`lightricks/ltx-2.5/text-to-video/pro`, `/fast`, i2v	$0.12 720p, $0.17 1080p	6/8/10s pro; up to 20s in other variants	1080p	Yes	No r2v endpoint seen	Claimed	VentureBeat, fal
BFL FLUX 3 video	Jul 23 2026 early access; fal Aug 4	`blackforestlabs/flux-3/text-to-video`, `/image-to-video`, `/keyframes-to-video`, `/first-last-frame-to-video`, `/draft` variants	$0.17 720p, $0.29 1080p	20s	1080p	Yes, multilingual dialogue	Image refs via i2v only	Agentic chaining	BFL blog, fal X
xAI Grok Imagine Video 1.5	GA Jun 17 2026	`xai/grok-imagine-video/v1.5/text-to-video`, `/image-to-video`, `/reference-to-video`, extend	$0.08 480p, $0.14 720p, $0.25 1080p	6s	1080p	Yes	Yes	No	x.ai, fal
Luma Ray 3.2	Jun 9 2026	`luma/agent/ray/v3.2/text-to-video`, `/image-to-video`	not captured	20s	1080p	Yes	16 keyframes (Luma app)	Keyframes	Luma
Runway Gen-4.5	Dec 1 2025	not on fal	n/a	10s	1080p HDR	Yes	Yes	No	Runway
OpenAI Sora 2	Sep 30 2025; product shut down Apr 26 2026	`fal-ai/sora-2/*` all deprecated	n/a	n/a	n/a	n/a	n/a	n/a	OpenAI
Vidu Q3 (Shengshu)	Jan 30 2026; R2V Apr 13	`fal-ai/vidu/q3/*`	not captured	16s	not stated	Yes	Yes	Smart cuts	PRNewswire
Pixverse v6 / C1	Mar 29 / Apr 8 2026	`fal-ai/pixverse/v6/`, `c1/`	not captured	not captured	not captured	Yes (switch)	C1 reference-to-video	multi-clip switch	fal catalog

Talking-head and lip-sync specialists on fal

fal-ai/bytedance/omnihuman/v1.5: image + audio to talking video, $0.16/s.
fal-ai/kling-video/lipsync/audio-to-video.
fal-ai/sync-lipsync/v3, fal-ai/sync-lipsync/v3/image-to-video (Jun 22 2026).
fal-ai/kling-video/create-voice (voice creation; wired in luxin).
Avatar endpoints: mirage-api/avatar-x/* (Aug 10 2026), fal-ai/heygen/*, fal-ai/ai-avatar/*.

Image models for reference sheets and keyframes

All active in the fal catalog on 2026-09-04: fal-ai/nano-banana-pro (+ /edit), fal-ai/nano-banana-2 (+ /edit), fal-ai/bytedance/seedream/v5/lite/text-to-image and /edit, fal-ai/bytedance/seedream/v4.5/*, fal-ai/gpt-image-1.5, fal-ai/flux-2-pro (+ /edit), fal-ai/gemini-3-pro-image-preview (+ /edit), fal-ai/z-image/*. Pricing not yet captured; add in Phase 3.

Community read on realistic talking humans, Aug 7 to Sep 4 2026

Single-creator tests, not reproduced. Ranking flips with content type and safety filters.

H3 best lip-sync in a same-script test; Kling 3.0 fails lip-sync and alters audio; Seedance 2.0/2.5 blocked by safety filters on that content (@hckinz Aug 7).
Seedance 2.5 skin and face realism "90 to 95%" vs 2.0; audio is the remaining gap; replies flag promo bias (@ChrisGPT Aug 8).
Wan 3.0 one-minute two-person podcast test held dual lip-sync and reactions (@MonetizationDon ~Aug 25).
Wan 3.0 wins on editing existing footage with lip-sync while keeping motion; ElevenLabs used for the voice track; reply says Wan audio quality below Seedance 2.5 and FLUX 3 (@fabianstelzer Aug 26).
Wan 3.0 fixed Kling 3.0 audio and Veo 3 lip-sync problems at about half Seedance 2.5 cost (@izbuilds Aug 27).
Seedance 2.5 lip-sync "perfect" on a sung clip; costly (@GuidedbyHumans Aug 31).
H3 best budget pick with strong lip-sync; Seedance 2.5 most realistic but lip-sync issues and expensive (@Kyle_Coghlan Sep 1).
On fal, H3 Max clip $0.20 vs H3 $1.30; H3 Max better acting, weaker voice quality; Gemini Omni "holds up" (@himanshuchanda Sep 2).
Four-way Seedance 2.5 / H3 / FLUX 3 / Wan 3.0 side-by-side, no single winner (@HBCoop_ Sep 2).

Net: Seedance 2.5 leads on face realism and 30s multi-shot consistency. Wan 3.0 and H3 / H3 Max lead on dialogue lip-sync per dollar. Kling 3.0 audio draws repeated complaints. Veo 3.1 is cited as strong but is the oldest and most expensive per second. Sora is gone.

Bake-off plan (Phase 5)

Same three shots on each: (1) solo confessional, 8s, one line of dialogue; (2) two-person scheme in the control room, 10s, two lines; (3) implied-action montage beat, 6s, no dialogue.

Wan 3.0 Prime reference-to-video, 720p draft then 1080p.
Seedance 2.5 reference-to-video, 720p only unless it wins outright.
MiniMax H3 Max reference-to-video and image-to-video, 768p (draft workhorse regardless of outcome). Not director: R5 found it is a WebRTC realtime stream with a 60 s minimum bill, not a batch multi-shot endpoint. Batch multi-shot on H3 Max is "CUT." inside the prompt.
Gemini Omni 1.1 Flash reference-to-video, 1080p, shot 1 only (hero confessional fallback).

Score per clip: identity match to reference sheet, lip-sync, acting naturalness, reality-TV look, filter refusals, cost. Winner per shot type, not overall.

Luxin overlap

Luxin's hosted runtime (registry 2026-08-12) wires 18 video models: Veo 3.1 fast/lite, Kling v3 standard, Seedance 2.0 fast, LTX 2.3 fast, Wan 2.7, Pixverse v6, Hailuo-02, Ray 3.2, ltx-video-13b. None of Wan 3.0, H3 Max, Omni Flash 1.1, Seedance 2.5 are wired. This project calls fal directly.

Not verified

Gemini Omni Flash duration options beyond 8s on fal. Kling v3 fal output resolution enum. alibaba/wan-3.0 non-Prime price. Vidu Q3 max resolution. Luma, Pixverse, Vidu fal prices. Image model prices. All X-post quality claims. Whether Seedance 2.5 filters reject "producer hands contestant a drink" style prompts (test in Phase 5).