Video Models · checked Oct 17, 2026
The video model race, scored.
Live rankings from Artificial Analysis and LMArena, real list prices per second, and the models ad teams are using. Every number links to its source.
Leaderboards
Each board keeps its own scale. Scores from different boards are never on one axis.
Text-to-Video (with audio): Artificial Analysis
Swipe the chart sideways for all 20 models →
Wan 3.0
Dreamina Seedance 2.5Seedance 2.0
Gemini Omni Flash 1.1
Grok Imagine Video 1.5HappyHorse-Wan2.7-HappyHorse-Kling 3.0 Omni 1080pPro
Kling 3.0 1080pProGrok Imagine Video 1.5 LiteElo (Kling 3.0 1080p Pro = 1000) · 20 models · top 5 coloured · axis does not start at zeroSource: artificialanalysis.ai ↗ · checked 2026-10-23
Image-to-Video: Artificial Analysis
Swipe the chart sideways for all 10 models →
Seedance 2.5Gemini Omni Flash 1.1Wan 3.0Seedance 2.0Grok Imagine Video 1.5HappyHorse-Elo · 10 models · top 5 coloured · axis does not start at zeroSource: artificialanalysis.ai ↗ · checked 2026-10-23
Text-to-Video: LMArena
Swipe the chart sideways for all 12 models →
Gemini Omni 1.1 FlashGemini Omni FlashGrok Imagine Video 1.5Seedance 2.0Wan 3.0Seedance 2.5
Muse VideoHappyHorse 1.0Veo 3.1Arena score · 12 models · top 5 coloured · axis does not start at zeroSource: lmarena.ai ↗ · board dated 2026-09-21, checked 2026-10-23
Image-to-Video: LMArena
Swipe the chart sideways for all 10 models →
Gemini Omni 1.1 FlashWan 3.0Seedance 2.5Seedance 2.0Gemini Omni FlashGrok Imagine Video 1.5HappyHorse 1.0Arena score · 10 models · top 5 coloured · axis does not start at zeroSource: lmarena.ai ↗ · board dated 2026-09-21, checked 2026-10-23
Image-to-Video: DesignArena (transcribed)
Swipe the chart sideways for all 20 models →
Wan 3.0Seedance 2.5Grok Imagine Video 1.5 PreviewSeedance 2.0Seedance 2.0 FastHappy Horse 1.1Grok ImagineSeedance 2.0 MiniHappy Horse 1.0Gemini Omni 1.1 FlashWan 2.7Gemini Omni FlashKling O3Elo (DesignArena) · 20 models · top 5 coloured · axis does not start at zeroSource: chart in @oneeyekeh post ↗ · board dated 2026-10-16, checked 2026-10-17
Transcribed from a vendor employee’s chart image; the live board couldn’t be parsed. The author later said HeyGen fell to #4 (post). Verify at designarena.ai.
On Oct 16, 2026, a HeyGen product lead posted HeyGen Video 1.0 at #1 on DesignArena image-to-video (1331). A day later he said it had been pushed down to #4. Every chart here shows its source and board date for that reason.
Score vs price
12 models
Cost per clip Estimate
cost = official list $/s × clip length (s); one generation, no retries, no promo, no tax. None = longer than the model's max single clip. Veo is capped at 8 s, so its 8 s column is shown. Real cost scales with retries; budget 2–3× for ad work.
| Model | Res. | $/s | 5 s | 8 s | 10 s | 15 s | Source |
|---|---|---|---|---|---|---|---|
| Wan 3.0 | 720p | $0.10 | $0.50 | $0.80 | $1.00 | $1.50 | modelstudio.alibabacloud.com ↗ |
| Wan 3.0 | 1080p | $0.20 | $1.00 | $1.60 | $2.00 | $3.00 | same as above |
| Seedance 2.5 | 720p | $0.23 | $1.16 | $1.85 | $2.31 | $3.47 | byteplus.com ↗ |
| Seedance 2.5 | 1080p | $0.57 | $2.84 | $4.55 | $5.69 | $8.54 | same as above |
| MiniMax H3 | 768p | $0.08 | $0.40 | $0.64 | $0.80 | $1.20 | platform.minimax.io ↗ |
| MiniMax H3 | 2K | $0.13 | $0.65 | $1.04 | $1.30 | $1.95 | same as above |
| HeyGen Video 1.0 | 768p | $0.03 | $0.15 | $0.24 | $0.30 | $0.45 | developers.heygen.com ↗ |
| HeyGen Video 1.0 | 1080p | $0.09 | $0.45 | $0.72 | $0.90 | $1.35 | same as above |
| Gemini Omni Flash 1.1 | 720p | $0.10 | $0.50 | $0.80 | $1.00 | over max | ai.google.dev ↗ |
| Veo 3.1 Lite | 1080p | $0.08 | $0.40 | $0.64 | over max | over max | ai.google.dev ↗ |
| Veo 3.1 Fast | 1080p | $0.12 | $0.60 | $0.96 | over max | over max | same as above |
| Veo 3.1 | 1080p | $0.40 | $2.00 | $3.20 | over max | over max | same as above |
| Grok Imagine Video 1.5 | default | $0.08 | $0.40 | $0.64 | $0.80 | $1.20 | docs.x.ai ↗ |
| Kling 3.0 (no audio) | 1080p | $0.11 | $0.56 | $0.90 | $1.12 | $1.68 | kling.ai ↗ |
| Kling 3.0 Turbo (audio) | 1080p | $0.14 | $0.70 | $1.12 | $1.40 | $2.10 | same as above |
| LTX-2.5 Fast | 1080p | $0.13 | $0.65 | $1.04 | $1.30 | $1.95 | ltx.io ↗ |
| LTX-2.5 Pro | 1080p | $0.17 | $0.85 | $1.36 | $1.70 | $2.55 | same as above |
| FLUX 3 Video | Full HD | $0.29 | $1.45 | $2.32 | $2.90 | $4.35 | bfl.ai ↗ |
HeyGen Video 1.0, Utopai X and Creatify Boreal-H3 are all post-trained on MiniMax H3. Base model for HeyGen Video 1.0, Utopai X and Creatify Boreal-H3 (each vendor says so).
Open source and ComfyUI
Models you can run yourself, with licence and the minimum VRAM people report.
| Model | Licence | ComfyUI | Minimum VRAM | Quantized |
|---|---|---|---|---|
| MiniMax H3 (open weights) released 2026-08-03 | minimax-h3-community-license-agreement; commercial use of local outputs needs a MiniMax commercial licence sold via Comfy | native templates (ComfyUI 0.30.0+); pruned INT8 ConvRot / bf16 / fp8 builds | Official: not stated Community report: 16 GB with Q4_0 (11.4 GB) or pruned GGUF + offload; ~26–32 GB peak with pruned INT8 at 864×480 5 s huggingface.co ↗ modelvram.com ↗ docs.comfy.org ↗ | INT8 ConvRot, fp8, GGUF (community), NVFP4 encoder |
| LTX-2.5 released 2026-08-11 | ltx-2.x-community-license-agreement (free under $10M ARR) | native, official LTX templates (T2V, I2V, FLF) | Official: 32 GB+ (LTX ComfyUI guide); lower-VRAM INT8 ConvRot / NVFP4 distilled files offered Community report: LTX marketing mentions 16 GB for quantized variants (search summary, unverified) ltx.io ↗ huggingface.co ↗ docs.ltx.io ↗ | INT8 ConvRot, NVFP4 |
| Wan 2.2 (TI2V-5B, T2V/I2V A14B) released 2025 (Wan 3.0 is API-only) | Apache 2.0 | native workflows + Kijai WanVideoWrapper + GGUF | Official: 5B fits ~8 GB with native offloading (ComfyUI docs) Community report: A14B fp8 ~14.5–19 GB at 720p/81 frames; Q4/Q5 GGUF ~10–15 GB with offload modelvram.com ↗ docs.comfy.org ↗ | fp8_scaled, GGUF |
| Wan 3.0 released 2026-08-24 | proprietary; no open weights | available in ComfyUI/Comfy Cloud as a hosted model, not local weights (no open weights per search summary) | Official: n/a (hosted) blog.comfy.org ↗ | — |
| HunyuanVideo 1.5 (8.3B) released 2025-11 | tencent-hunyuan-community | native T2V/I2V templates | Official: designed for 24 GB consumer GPUs (ComfyUI docs); README cites a 14 GB minimum with offloading (search summary) Community report: 12 GB OOM at 120 frames github.com ↗ huggingface.co ↗ docs.comfy.org ↗ | fp8 text encoder |
| Mochi 1 preview (Genmo) released 2024-10 | Apache 2.0 | native (model card: ComfyUI can run it on <20 GB) | Official: 42 GB full / 22 GB bf16 (diffusers); <20 GB in ComfyUI per model card huggingface.co ↗ | bf16, fp8 (ComfyUI) |
| CogVideoX 1.5-5B (Zhipu) released 2024-11 | CogVideoX licence ('other') | community wrapper (kijai ComfyUI-CogVideoXWrapper); not verified in Comfy docs | Official: diffusers BF16 from 10 GB, INT8 from 7 GB (with CPU offload) huggingface.co ↗ | INT8 torchao, fp8 |
| Cosmos Predict2 2B (NVIDIA) released 2025 | NVIDIA open model licence (not checked) | native Video2World template | Official: ~16 GB for 2B (ComfyUI docs) docs.comfy.org ↗ | — |
Discontinued and watchlist
Shutdown announced 2026-03-24; app/web ended 2026-04-26; API removal scheduled 2026-09-24
apnews.com ↗ developers.openai.com ↗ checked 2026-10-17previewed 2026-07-07, not generally available; LMArena T2V #9 (1456)
ai.meta.com ↗ lmarena.ai ↗AA T2V #11 (1055), Higgsfield $0.14/s list
artificialanalysis.ai ↗ higgsfield.ai ↗lower on the boards; not profiled
artificialanalysis.ai ↗ lmarena.ai ↗Sources and gaps
- Artificial Analysis Video Arena: Text-to-Video v2.1 ↗ Elo anchored so Kling 3.0 1080p Pro = 1000; blind votes from validated private evaluators
- Artificial Analysis: Image-to-Video ↗
- LMArena Text-to-Video ↗ · board 2026-09-21 · 718,577 votes
- LMArena Image-to-Video ↗ · board 2026-09-21 · 2,081,953 votes
- DesignArena Video / Image-to-Video ↗ client-rendered; scores taken only from the seed post's chart image
- VBench ↗ academic automatic benchmark; no current 2026 leaderboard for these models found
- No verified official spend or ROI data for any brand ad campaign; ad use cases are spec, contest or vendor items
- FLUX 3, Utopai and Boreal prices come from secondary sources
- DesignArena scores come only from the vendor's chart image
- No current VBench scores for 2026 models
- Kling 4.0 launch date unverified
- Seedance 2.5 and MiniMax per-second prices come via summaries of official pages that weren't fully parsed
- Open-model VRAM minimums are mostly community figures; few vendors state one
- Wan 3.0 is not open-weights; ComfyUI access is hosted
FAQ
Which AI video model is best right now?
On Artificial Analysis text-to-video it is Wan 3.0 (1175); on LMArena text-to-video it is Gemini Omni 1.1 Flash (1516, board dated 2026-09-21). Checked 2026-10-17.
How much does an AI video clip cost?
Estimate: official list $ per second times clip length. A 15-second 720p Wan 3.0 clip is about $1.50 for one generation; budget 2–3x for retries.
Is Sora 2 still available?
Shutdown announced 2026-03-24; app/web ended 2026-04-26; API removal scheduled 2026-09-24