Reference model
Bottom line
Seedance is the complete model; HappyHorse is the quality test; Veo is the API ladder.
Use Seedance 2.0 when you need a single model to understand references across text, image, audio, and video and produce native audiovisual, multi-shot output. Use HappyHorse-1.0 when you want to test the current highest no-audio blind-preference leader through fal. Use Veo 3.1 when official Google API availability, Lite/Fast/Standard pricing tiers, vertical output, and 4K upscaling matter more than one benchmark row.
Scope
Only 2026 releases and major 2026 updates are in the main table.
| Included model | 2026 signal | Why it belongs |
|---|---|---|
| HappyHorse-1.0 | April 2026 fal/partner launch. | Current quality leader in the no-audio benchmark row plus direct fal access. |
| Seedance 2.0 | ByteDance launch dated 2026-02-12. | The target model and the strongest multimodal audiovisual package in this pass. |
| Veo 3.1 / Veo 3.1 Lite | Google January 2026 update and March 2026 Lite release. | Official API availability, vertical output, 1080p/4K options, and lowest official Veo tier. |
| Kling 3.0 / Video 3.0 Omni | Kuaishou launch dated 2026-02-05. | Best first-party storyboarding and production-control framing. |
| Vidu Q3 Pro | Current 2026 Q3 API/pricing docs and benchmark row. | Transparent 1-16 second official per-second pricing. |
| PixVerse V6 | PixVerse launch dated 2026-03-30. | Native-audio, multi-shot, agentic/social workflow release. |
| Grok Imagine Video | xAI video docs updated in 2026 and current benchmark row. | Low-cost short-form generation/editing option. |
| Luma Ray3.14 | Luma launch dated 2026-01-26. | Native 1080p, faster/cheaper iteration, and Modify Video up to 18 seconds. |
| SkyReels V4 | February/March 2026 technical report. | Unified video-audio generation, inpainting, editing, 1080p, 32 FPS, and 15-second paper claims. |
| LTX-2.3 | March 2026 release and API changelog. | Open/local pipeline with published per-second API prices. |
Not in scope
Important leaders held out by the 2026-only rule
| Model | Why it is not a main row | How to use it anyway |
|---|---|---|
| Runway Gen-4.5 | Artificial Analysis lists a December 2025 release. | Still worth benchmarking against top 2026 rows for realism and production workflows. |
| Sora 2 / Sora 2 Pro | Useful market reference, but not a 2026 model release in the cited source set. | Use as a comparison point for long single takes, character IDs, and platform access. |
| Hailuo 02 / 2.3 | Visible benchmark rows point to 2025 releases. | Keep as a lower-cost practical baseline outside this strict guide. |
| Pika 2.5 | Artificial Analysis lists a November 2025 release. | Still relevant for consumer effects and app workflows. |
Quality
Quality rank is real signal, but workflow fit decides the winner.
| Model | Quality read | Best use-case | Main caveat |
|---|---|---|---|
| HappyHorse-1.0 | Current no-audio Artificial Analysis leader and 1080p fal model. | Ads, product clips, e-commerce, short social spots, API quality trials. | Newer ecosystem and less long-run reliability evidence. |
| Seedance 2.0 | Strongest complete audio-video plus multimodal-reference package. | Multi-shot audiovisual scenes, brand references, storyboards, product campaigns. | Access, regional availability, provider pricing, and rights concerns need review. |
| Veo 3.1 / Lite | Best official Google developer ladder in this set. | Developer products, vertical mobile video, Flow/API workflows, 4K upscale needs. | Standard tier is expensive; Lite and Fast trade off quality. |
| Kling 3.0 / Omni | Strong storyboarding, narrative control, and character/reference consistency. | Director-style prompts, product visualization, ads, multi-shot planning. | Provider/credit pricing varies; benchmark rows trail HappyHorse and Seedance. |
| Vidu Q3 Pro | Practical API model with clear pricing and 1-16 second output. | Volume generation, reference-to-video, cost-aware API experiments. | Not the top quality leader. |
| PixVerse V6 | Good current social/agentic workflow release with native audio. | Fast branded social clips, agentic creative pipelines, short multi-shot ads. | Less source-backed detail on public API cost and lower absolute quality signal. |
| Grok Imagine Video | Low-cost current benchmark entry. | Cheap drafts, fast ideation, social-format video, image/video edits. | 480p/720p limit, moderation behavior, and brand-safety review matter. |
| Luma Ray3.14 | Professional edit/cinematic workflow, now native 1080p. | Modify Video, polished image/video-to-video, HDR-adjacent workflows. | Not as native-audio-first as Seedance, Kling, HappyHorse, or Veo. |
| SkyReels V4 | Strong technical-paper signal for unified audio-video plus editing. | Research, open-model exploration, video-audio inpainting/editing experiments. | Commercial access and dollar pricing were not cleanly confirmed. |
| LTX-2.3 | Best open/local and transparent-cost option, behind closed leaders on broad preference rankings. | Local generation, reproducible pipelines, open weights, vertical 1080p. | Quality ceiling is lower than the strongest closed 2026 models. |
Specs
Resolution, duration, inputs, and audio support
| Model | Resolution / duration | Inputs and control | Audio |
|---|---|---|---|
| HappyHorse-1.0 | 720p or 1080p; 3-15 seconds on fal. | Text-to-video, image-to-video, reference-to-video, video edit; common social aspect ratios. | Native synchronized audio. |
| Seedance 2.0 | Paper: native 480p/720p; 4-15 seconds. Some providers expose 1080p routes. | Text, image, audio, video; up to 9 images, 3 videos, and 3 audio references in ByteDance/fal descriptions. | Native audio-video generation with dual-channel/stereo-style sound design claims. |
| Veo 3.1 | 720p/1080p plus 4K output path; vertical 9:16 support in 2026 update. | Text/image-to-video, Ingredients to Video, Flow/API/Vertex workflows. | Google pricing page labels video-with-audio as default. |
| Veo 3.1 Lite | 720p/1080p; 4s, 6s, or 8s durations. | Text-to-video and image-to-video; landscape and portrait framing. | Video with audio by default in pricing table. |
| Kling 3.0 / Omni | 720p/1080p up to 15 seconds. | Text/image/audio/video inputs, reference videos/images, multi-shot storyboard, in-video editing. | Native audio across multiple languages, accents, dialects, and voice-control mode. |
| Vidu Q3 Pro | 540p/720p/1080p; 1-16 seconds. | Text-to-video, image-to-video, start-end-to-video, reference-to-video. | Audio/video features vary by endpoint; price table is video-generation focused. |
| PixVerse V6 | Benchmark row shows 720x1280; platform emphasizes multi-shot short films. | Camera control, character performance, agentic/CLI workflows, multi-shot prompts. | Native audio in the V6 release positioning. |
| Grok Imagine Video | 480p or 720p; max generation duration 15 seconds. | Text-to-video, image-to-video, edits, extensions through xAI video endpoints. | Integrated video-generation workflow; confirm exact audio behavior by endpoint. |
| Luma Ray3.14 | Native 1080p; Modify Video up to 18 seconds. | Text-to-video, image-to-video, video-to-video, Modify Video; HDR limitations noted. | Audio is not the central Ray3.14 advantage; Luma has separate add-audio/API pricing surfaces. |
| SkyReels V4 | Paper: up to 1080p, 32 FPS, 15 seconds. | Text, images, video clips, masks, audio references; generation, inpainting, editing. | Joint video-audio generation in a dual-stream architecture. |
| LTX-2.3 | 1080p, 1440p, and 4K API tiers; native portrait up to 1080x1920. | Text-to-video, image-to-video, audio-to-video, retake, extend, local desktop workflows. | Synchronized audio model and audio-to-video endpoint. |
Costs
Per-second and per-minute price signals
| Model | Published price signal | Normalized minute cost | Pricing caveat |
|---|---|---|---|
| HappyHorse-1.0 | fal: $0.14/s at 720p; $0.28/s at 1080p. | $8.40-$16.80/min. | Provider route; first-party Alibaba/ATH pricing may differ. |
| Seedance 2.0 | fal: $0.2419/s fast and $0.3034/s standard at 720p; TechNode reported Volcengine around 1 RMB/s for pure generation. | $14.51-$18.20/min on fal; Volcengine estimate depends on exchange rate and access. | ByteDance/Dreamina access and provider routes differ; 1080p pricing varies. |
| Veo 3.1 Standard | Google Gemini API: $0.40/s at 720p/1080p; $0.60/s at 4K. | $24.00-$36.00/min. | Paid tier only; preview-model behavior may change. |
| Veo 3.1 Fast | Google Gemini API: $0.10/s at 720p; $0.12/s at 1080p; $0.30/s at 4K. | $6.00-$18.00/min. | Lower cost than Standard, with quality/latency tradeoffs. |
| Veo 3.1 Lite | Google Gemini API: $0.05/s at 720p; $0.08/s at 1080p. | $3.00-$4.80/min. | No 4K output; intended as the lower-cost developer tier. |
| Kling 3.0 / Omni | Official docs: 6-12 credits/s depending resolution/audio, plus 2 credits/s voice control; Artificial Analysis rows show $10.08-$13.44/min. | $10.08-$13.44/min in benchmark row. | Credits do not convert cleanly without plan assumptions. |
| Vidu Q3 Pro | Official API: $0.05/s 540p, $0.125/s 720p, $0.15/s 1080p; off-peak $0.025-$0.075/s. | $3.00-$9.00/min regular; $1.50-$4.50/min off-peak. | Off-peak pricing may not fit latency-sensitive production. |
| PixVerse V6 | Artificial Analysis: $5.40/min. | $5.40/min. | Use current provider docs before buying; official launch post does not provide a dollar table. |
| Grok Imagine Video | Artificial Analysis: $4.20/min; xAI docs say duration and resolution affect per-second cost. | $4.20/min benchmark row. | Official model-page pricing should be rechecked before routing volume. |
| Luma Ray3.14 | Luma pricing: 10 credits/s at 720p and 20 credits/s at 1080p for T2V/I2V; HDR 40/80 credits/s. | Credit-based, not converted to dollars here. | Dollar cost depends on plan credits and Luma account tier. |
| SkyReels V4 | No clean first-party dollar price found. | Not normalized. | Treat as research/watchlist until access and billing are clear. |
| LTX-2.3 | API: $0.06-$0.32/s for T2V/I2V Fast/Pro across 1080p, 1440p, and 4K; $0.10/s for A2V/retake/extend. | $3.60-$19.20/min for T2V/I2V tiers; $6.00/min for A2V/retake/extend. | Local generation shifts cost from API spend to hardware/time. |
Use-cases
The shortest decision table
| Need | Best pick | Why |
|---|---|---|
| Best single quality experiment | HappyHorse-1.0 | It has the strongest current no-audio blind-preference signal and direct fal pricing. |
| Best all-around audiovisual generation | Seedance 2.0 | It combines multimodal references, native audio-video output, multi-shot generation, editing, and extension. |
| Best official API price ladder | Veo 3.1 family | Google publishes Standard, Fast, and Lite per-second prices with 720p, 1080p, and selected 4K tiers. |
| Best storyboarding and director control | Kling 3.0 Omni | The first-party launch emphasizes multi-shot storyboard control and reference consistency. |
| Best transparent volume pricing | Vidu Q3 Pro | The official pricing table exposes per-second regular and off-peak rates across 540p, 720p, and 1080p. |
| Best edit/cinematic polish | Luma Ray3.14 | Ray3.14 focuses on native 1080p, faster iteration, cleaner results, and Modify Video up to 18 seconds. |
| Best open/local workflow | LTX-2.3 | It ships open weights, an open desktop editor, native portrait support, and clear API pricing. |
| Best research watchlist | SkyReels V4 | Its paper describes unified video-audio generation, inpainting, editing, and 1080p/32 FPS/15s support. |
Watch-outs
The details that change model choice
- Access matters as much as quality: a top model with no stable API may be worse for production than a slightly weaker model with reliable billing and SDKs.
- Native 1080p, 1080p upscale, and provider-output 1080p are not equivalent. Check whether the source says native generation or post-generation upscale.
- Audio is not one feature. Dialogue, Foley, ambient sound, music, lip-sync, audio references, and add-on audio are separate capabilities.
- Credits are not dollars until the plan or bundle conversion is known. Do not compare credit rates across services directly.
- Quality benchmarks are prompt-distribution dependent. Run a matched prompt set for the actual job before committing spend.
- Identity, copyright, and brand-reference use needs policy review, especially for Seedance-style reference workflows and real-person inputs.
Receipts
How this guide was built
This guide was built from first-party model pages, official release notes, API pricing pages, technical papers, fal model/provider pages, and Artificial Analysis benchmark rows checked on 2026-05-03. The strict inclusion rule is 2026 release or major 2026 model-family update only; late-2025 leaders are named but held out of the main comparison.
Receipts
Claim table
| Where | Claim | Confidence | Source | Caveat |
|---|---|---|---|---|
| Scope | ByteDance Seed officially launched Seedance 2.0 on 2026-02-12. | High | ByteDance Seed | Regional product availability and naming may differ across Dreamina, Seed, Volcengine, and provider routes. |
| Specs | Seedance 2.0 supports text, image, audio, and video inputs and can use multiple image, video, and audio references. | High | ByteDance Seed | Exact limits can differ by endpoint and provider wrapper. |
| Specs | The Seedance 2.0 technical paper describes 4-15 second direct audio-video generation and native 480p/720p output. | High | arXiv | Provider output modes may add higher-resolution routes or upscaling. |
| Costs | fal listed Seedance 2.0 at $0.2419/s fast and $0.3034/s standard for 720p text-to-video. | Medium | fal.ai | This is a provider price, not necessarily ByteDance first-party pricing. |
| Specs | fal lists HappyHorse-1.0 text-to-video as 720p/1080p, 3-15 seconds, common social aspect ratios, and native synced audio. | High | fal.ai | fal page details can change as model access matures. |
| Costs | fal lists HappyHorse-1.0 at $0.14/s for 720p and $0.28/s for 1080p. | High | fal.ai | Provider-specific pricing may differ from Alibaba first-party or other gateway prices. |
| Quality | Artificial Analysis listed HappyHorse-1.0, Seedance 2.0, Kling 3.0, Grok Imagine Video, Vidu Q3 Pro, PixVerse V6, and LTX-2.3 in current 2026 benchmark rows. | Medium | Artificial Analysis | Leaderboard ranks and Elo scores move as additional blind votes are collected. |
| Scope | Kuaishou announced Kling AI 3.0 on 2026-02-05 with Video 3.0, Video 3.0 Omni, Image 3.0, and Image 3.0 Omni. | High | Kuaishou Technology | Initial access was described as Ultra-subscriber early access, with broader rollout timing subject to change. |
| Specs | Kling 3.0 supports up to 15-second generation, native audio across languages, multimodal workflows, reference consistency, and multi-shot storyboard control. | High | Kuaishou Technology | Detailed settings and credit costs depend on Video 3.0 versus Omni and account access. |
| Scope | Google's January 2026 Veo 3.1 update added native vertical Ingredients to Video and 1080p/4K upscale availability through Flow/API/Vertex paths. | High | Availability and limits can differ between Gemini app, Flow, Gemini API, Vertex AI, YouTube, and Google Vids. | |
| Specs | Google introduced Veo 3.1 Lite on 2026-03-31 as its most cost-effective video model, supporting text-to-video and image-to-video at 720p/1080p for 4s, 6s, or 8s durations. | High | Lite is explicitly a lower-cost tier and should be tested against quality requirements. | |
| Costs | Google Gemini API lists Veo 3.1 Standard, Fast, and Lite per-second prices from $0.05/s to $0.60/s depending tier and resolution. | High | Google AI for Developers | Preview model pricing and rate limits may change before stable release. |
| Costs | Vidu Q3 Pro lists official API pricing from $0.05/s at 540p to $0.15/s at 1080p, with lower off-peak rates. | High | Vidu | Off-peak rates may not be appropriate for latency-sensitive jobs. |
| Quality | PixVerse launched V6 on 2026-03-30 with improved camera control, character performance, multi-shot video, and native audio positioning. | High | PixVerse | The official launch post is feature-focused and does not provide the full pricing table. |
| Specs | xAI documentation lists Grok Imagine Video generation limits including maximum 15-second generation duration and 480p/720p resolutions. | High | xAI Docs | Exact pricing belongs on the current model page and should be rechecked before routing volume. |
| Quality | Luma launched Ray3.14 on 2026-01-26 with native 1080p, 4x faster 720p generation, 3x cheaper 720p generation versus Ray3, and Modify Video up to 18 seconds. | High | Luma | The launch post compares to Ray3 and does not make Ray3.14 a native-audio-first model. |
| Costs | Luma's pricing page lists Ray3.14 generation costs in credits per second, including 10 credits/s at 720p and 20 credits/s at 1080p for text-to-video/image-to-video. | High | Luma | Credit rates are not directly comparable to dollar per-second pricing without plan conversion. |
| Quality | LTX-2.3 shipped in March 2026 with an open-source desktop editor, open weights, improved details, better prompt understanding, improved image-to-video, cleaner audio, and native portrait support. | High | LTX / Lightricks | Local output quality depends on workflow, hardware, and configuration. |
| Costs | LTX docs list LTX-2.3 Fast/Pro text-to-video and image-to-video prices from $0.06/s to $0.32/s depending resolution and model tier. | High | LTX Documentation | Audio-to-video, retake, extend, and HDR endpoints have separate rates. |
| Specs | The SkyReels V4 paper describes a unified video-audio model for generation, inpainting, and editing with up to 1080p, 32 FPS, and 15-second output. | Medium | arXiv | Paper claims do not substitute for stable commercial access, pricing, or independent production testing. |
| Not in scope | Runway Gen-4.5, Sora 2, Hailuo 2.3, and Pika 2.5 are excluded from the main table because the cited rows or release evidence point outside the strict 2026-release rule. | Medium | Artificial Analysis | They may still be relevant market references outside this guide's strict inclusion boundary. |
Sources