The Complete AI Video Model Comparison: Runway, Veo, Gemini Omni, Kling, Seedance, Wan, and LTX-2.3

Artificial Analysis rankings re-checked October 8, 2026, plus specs and per-second pricing for the major AI video models (specs and prices last verified July 18, 2026, except where noted).

Seven models, seven different bets on what AI video should be. Abstract editorial illustration by RizzGen.

The AI video field moved fast between February and October 2026. Runway Gen-4.5 launched claiming the top benchmark spot, then disappeared from the live leaderboard as newer entrants arrived. Google folded Veo into a broader multimodal model called Gemini Omni. ByteDance shipped native 4K on Seedance 2.0, then followed it with Seedance 2.5. Alibaba's Wan and Kuaishou's Kling both jumped a full version, and Alibaba's Wan 3.0 now leads the text-to-video board. OpenAI's Sora is not part of this comparison: the consumer app was discontinued on April 26, 2026, and OpenAI scheduled the API to shut down on September 24, 2026.

Every number below is sourced to an official vendor page, official API documentation, or the live Artificial Analysis Text-to-Video Arena — not to a marketing tweet repeated across a dozen SEO blogs. Where sources disagree or a model is too new to have public data, we say so instead of guessing.

The Leaderboard Was Reset: What Changed

Artificial Analysis rebuilt its text-to-video benchmark as AA-Video-T2V v2.0. The new version uses a refreshed prompt set, judges every model at 1080p (or the closest resolution the provider supports), evaluates videos with synchronized audio, and discarded every vote cast under the earlier methodology. Elo was recalculated from scratch, with Kling 3.0 (1080p Pro) anchored at 1,000. Two consequences matter if you have read this comparison, or any other, before: scores from earlier versions are not comparable with today's numbers, and the order changed. Gemini Omni Flash, first on the board in July, is now tied for seventh–eighth, while Seedance 2.5, which had no arena data in July, ranks third.

Several models now rank at or near the top that this page has not yet profiled. We have not verified their specifications or pricing, so they are listed here for reference only and are not in the comparison table or the compare tool:

Interactive Comparison

Filter the table by capability, or click any column header to sort.

Compare Two Models Head-to-Head

Pick any two models for a direct spec-by-spec comparison. Lower price and higher Elo are highlighted — everything else is a trade-off, not a win.

1. Runway Gen-4.5 (December 2025)

Runway released Gen-4.5 on December 1, 2025 and announced it had scored 1,247 Elo on the Artificial Analysis Text-to-Video benchmark, claiming the top spot at the time. That claim is the source of the "#1 AI video model" line you'll see repeated across dozens of comparison articles, including the earlier version of this one.

It's no longer verifiable as current. As of October 8, 2026, Runway Gen-4.5 still does not appear on Artificial Analysis's live Text-to-Video or Image-to-Video leaderboards. We can't tell you whether that's because Runway stopped submitting the model for evaluation, the listing was retired, or something else; we can only tell you what the live board shows today. The 1,247 figure also sits on a scale that no longer exists: Artificial Analysis rebuilt its text-to-video benchmark as v2.0 and discarded every earlier vote. Treat 1,247 as a historical, self-reported claim from launch day, not a current ranking.

Key Specifications

• Max Duration: Runway's API reference accepts 2–10 seconds; 5s and 10s are the commonly offered presets
• Resolution: 1080p is what integrators consistently report; Runway has not published an official resolution spec sheet for Gen-4.5
• Native Audio: Added in the December 2025 update — synchronized dialogue, ambient sound, and music
• Pricing: Runway's own docs list two different per-second rates in different places (12 and 25 credits/second, i.e. roughly $0.12–$0.25/s) — likely tier- or resolution-dependent; we couldn't find a single official number that reconciles both
• Architecture: Diffusion transformer, trained and served on NVIDIA Hopper/Blackwell GPUs

Strengths

Character Consistency: Still widely regarded as best-in-class for maintaining character appearance across scenes without fine-tuning, using reference images to lock subjects across lighting and camera changes.

Prompt Adherence: Strong instruction-following on camera movement language ("dolly zoom," "whip pan," "tracking shot").

Limitations

• No longer independently benchmarked on the leaderboard it built its launch marketing around
• Runway itself acknowledges causal reasoning errors and object permanence issues in complex scenes
• Thin official documentation on exact resolution and duration limits, which makes it harder to plan production pipelines around

Best For

Teams that already value Runway's character-locking workflow and don't need the highest current benchmark score to justify the choice. If a top Artificial Analysis ranking specifically is the deciding factor, start from the current v2.0 board, where Wan 3.0, Utopai X and Seedance 2.5 lead.

2. Google: Veo 3.1 and Gemini Omni Flash

Google now ships two distinct video-generation products, and conflating them is the most common mistake in other 2026 comparison articles. Veo 3.1 is Google's specialized cinematic video model, released October 14, 2025, and it remains the recommended "video model baseline" in the official Gemini API and Vertex AI documentation — its model ID (veo-3.1-generate-preview) is still live. Gemini Omni Flash is a newer, separate multimodal model announced May 19, 2026 at Google I/O that replaced Veo inside the consumer Gemini app, but has not replaced Veo 3.1 in the developer API, where Google says broader access is still rolling out.

Best For

Choose Veo 3.1 if you need a stable, documented API with enterprise/Vertex AI deployment today. Choose Gemini Omni Flash if conversational, iterative editing matters more to your workflow than topping a raw-quality bar — once its API is broadly available.

3. LTX-2.3 (March 2026)

Lightricks' LTX-2.3 is an open-source, 22-billion-parameter diffusion transformer (Lightricks' newer LTX-2.5 now appears on the Arena board; see the Elo note below) that generates synchronized audio and video in a single forward pass.

Key Specifications

• Max Duration: Up to 20 seconds, single pass
• Resolution: Up to 4K at 50fps
• Native Audio: Yes, generated simultaneously with video; this version improved audio quality via filtered training data and an upgraded vocoder
• License: Open weights, four checkpoint variants (dev, distilled, fast, pro) — the distilled variant runs in as few as 8 denoising steps
• Hardware: Runs locally via LTX Desktop on consumer GPUs, or via the Lightricks API
• Pricing: $0.06/s (1080p fast) to $0.32/s (4K pro) on Lightricks' own API, effective April 1, 2026; comparable third-party hosts (e.g. fal.ai) price similarly by resolution tier
• Artificial Analysis Elo: 892 (Pro) and 855 (Fast) on the v2.0 text-to-video board; 949 (Pro) on image-to-video (v1.0). The newer LTX-2.5 scores higher (946 Fast, 943 Pro on the text-to-video board); this page has not yet verified its specs or pricing

Strengths

True open weights: Full model access, fine-tunable, deployable on-premise with no API dependency.

Only model with native 4K as a default capability across its full duration range, not a separately-priced add-on tier bolted onto a lower-res base model.

Limitations

• Sits well below the top of the current Arena text-to-video board (892 Elo for Pro versus 1,156 for the leader) and has been outscored by the newer LTX-2.5
• Self-hosting the full 22B model requires real GPU budget even though the API is cheap

Best For

Developers and studios that need on-premise deployment, fine-tuning, or the cheapest 4K output on the market, and can tolerate a quality gap versus the top proprietary models.

4. Wan 2.7 (April 2026)

Alibaba's Tongyi Lab shipped the full Wan 2.7 suite between April 1–6, 2026 — four models covering text-to-video, image-to-video, reference-to-video (with voice cloning), and instruction-based editing, all under Apache 2.0.

Key Specifications

• Max Duration: 2–15 seconds for text/image generation; 2–10 seconds for reference and editing workflows
• Resolution: 720p or 1080p at 30fps
• Native Audio: Not confirmed in Wan 2.7's public documentation. Its predecessor, Wan 2.6, supported audio-visual sync — if that carried forward, Alibaba hasn't stated it plainly in the materials we could find, so we're not asserting it here
• Key new feature: First-and-last-frame control (specify both ends, the model fills the middle) and natural-language video editing across character actions, dialogue, style, and camera technique
• License: Apache 2.0, API from $0.10/second
• Artificial Analysis Elo: 1,030 (Wan2.7-260612, text-to-video v2.0, rank 13); 1,077 (image-to-video v1.0). Alibaba's newer Wan 3.0 now leads the text-to-video board at 1,156

Strengths

Genuinely open: Apache 2.0 licensing with full model access. It ranked third on the text-to-video board in July; after the benchmark reset and the arrival of newer entrants, including its own successor Wan 3.0, it now sits at rank 13.

First/last-frame control is a precision tool most proprietary models don't expose directly.

Limitations

• Caps at 1080p — no 4K tier
• Audio capability unclear from public documentation

Best For

Developers who want an open, self-hostable Apache-licensed model, especially for precise start/end-frame control, and who can accept a mid-table leaderboard position.

5. ByteDance: Seedance 2.0 and Seedance 2.5

Best For

Seedance 2.0 for the best-documented combination of visual quality, native 4K, and multimodal reference control. Seedance 2.5 now outscores it on the text-to-video board (1,143 versus 1,116) and runs to 30 seconds, but it has no official per-second rate and third-party reports say it ships at 720p maximum, so confirm availability, resolution and pricing with your provider before committing a production pipeline to it.

6. Kling 3.0 (February 2026)

Kuaishou announced the Kling AI 3.0 family — Video 3.0, Video 3.0 Omni, Image 3.0, and Image 3.0 Omni — on February 4, 2026. It's a distinct, later release from "Avatar 2.0," which Kuaishou shipped in November 2025 specifically for 5-minute continuous avatar generation; the two are often conflated in other comparison articles, but they're separate features with separate duration limits.

Key Specifications

• Max Duration: 15 seconds for general video generation (up from 10s on the prior version). The separate Avatar mode, introduced in November 2025, supports up to 5 continuous minutes for talking-head/avatar content specifically — not the general text-to-video path
• Resolution: Native 4K at up to 60fps (up from 1080p/48fps)
• Native Audio: Yes, on the base model — Kuaishou's launch materials list native audio generation in English, Chinese, Japanese, Korean, and Spanish, including regional accents and dialects, plus multi-character dialogue scenes where each character can speak a different language. Kling 3.0 Omni is a separate variant, but its differentiator is reference-based voice and character consistency (upload a reference video and it locks visual and vocal traits across scenes) plus multi-shot storyboarding — not audio generation itself, which the base model already has
• Pricing: $0.084/s (standard, no video input) to $0.168/s (Pro, with video input) on the official per-second API; consumer plans are credit-based (Free: 66 credits/day; Pro: $29.99/mo for 3,000 credits; Ultra: $59.99/mo for 8,000 credits plus 4K/60fps access)
• Artificial Analysis Elo: 1,000 (1080p Pro, the fixed anchor of the v2.0 text-to-video board) and 1,015 (Omni 1080p Pro); 1,055 (1080p Pro) and 1,044 (Omni 1080p Pro) on image-to-video v1.0

Strengths

Avatar Performance: The dedicated Avatar mode's 5-minute continuous generation remains unmatched among the models covered here for long-form talking-head content.

Native 4K/60fps is now standard on the flagship tier, not an add-on.

Multi-language native audio on the base model, including per-character language control within the same multi-character scene — a level of dialogue control not documented on any other model in this comparison.

Limitations

• The Omni variant — reference-based voice/character consistency and multi-shot storyboarding — scores slightly higher than the standard Pro tier on text-to-video (1,015 versus 1,000) but lower on image-to-video (1,044 versus 1,055), so the gap between them is small
• Credit-based consumer pricing makes cost comparison across plans harder than a flat per-second rate

Best For

Virtual-presenter and long-form avatar content specifically (via Avatar mode), or general 4K/60fps generation on a lower per-second cost than Veo or Runway.

Use Case Recommendations

Choose Runway Gen-4.5 If...

• Character consistency across multiple scenes is your top priority
• You value Runway's editing ecosystem over chasing the current top benchmark score

Choose Veo 3.1 If...

• You need a stable, documented, enterprise-ready API today
• You're already on Google Cloud/Vertex AI infrastructure
• You need long chained sequences (140s+) via scene extension

Choose Gemini Omni Flash If...

• Multi-turn conversational editing matters more than topping a raw-quality benchmark
• You can work within the Gemini app today and wait for broader API access

Choose LTX-2.3 If...

• You require on-premise deployment or fine-tuning
• You want native 4K without paying a proprietary 4K surcharge
• Budget is the binding constraint ($0.06/s fast tier)

Choose Wan 2.7 If...

• You want an Apache-licensed, self-hostable model and can accept a mid-table leaderboard position (1,030 Elo, rank 13)
• First/last-frame control is important to your workflow

Choose Seedance 2.0 If...

• You want the best currently-available combination of visual quality and native 4K
• You're working through ByteDance's consumer apps rather than waiting on the standalone API

Choose Seedance 2.5 If...

• You need 30-second native clips or local (partial-scene) editing, and can confirm availability, resolution and pricing with your provider first

Choose Kling 3.0 If...

• You need 5-minute continuous avatar/talking-head generation
• You want native 4K/60fps at a lower per-second rate than Veo or Runway

How RizzGen Uses These Models

RizzGen doesn't lock you into one video model. Our routing layer evaluates every model covered in this comparison against what your shot actually needs — resolution, audio, motion complexity, reference count, budget — and automatically selects the best-fit engine for that specific generation. We then apply RizzGen's own scene-based character locking and multi-scene consistency layer on top, so switching which underlying model handles a shot doesn't mean re-learning a new prompt syntax or losing character continuity between scenes.

We don't replace these models. We make the choice between them automatic, and make them controllable for production workflows.

FAQ

Sources & Methodology

Artificial Analysis rankings were checked directly against the live boards on October 8, 2026 (AA-Video-T2V v2.0 and AA-Video-I2V v1.0). Specifications and prices were checked against official vendor documentation on July 18, 2026, apart from Seedance 2.5, which was re-checked on October 8, 2026 using ByteDance-ecosystem coverage and reseller pricing pages because ByteDance has not published an official per-second rate. Where sources disagreed or a figure wasn't officially published, we said so in the text rather than presenting a single unverified number as fact. Key primary sources used:

• Artificial Analysis — Text-to-Video Arena Leaderboard
• Artificial Analysis — Image-to-Video Arena Leaderboard
• Artificial Analysis — Video Generation Benchmarking Methodology (v2.0)
• OpenAI — What to know about the Sora discontinuation
• Runway — Introducing Gen-4.5 and Runway API Pricing Docs
• Google — Video Generation in the Gemini API (Veo 3.1) and Gemini Omni Overview
• Lightricks — LTX-2.3 on Hugging Face
• Kuaishou Investor Relations — Kling AI 3.0 Launch
• ByteDance/Volcano Engine Seedance 2.0 and 2.5 pricing and rollout coverage (BigGo Finance, BytePlus resource pages)
• Seedance 2.5 reseller pricing and availability, October 2026: OpenRouter, CellCog, Apiframe
• Alibaba Wan 2.7 official model release coverage (April 2026)

This is a fast-moving space — pricing, availability, and benchmark rankings can change within weeks. If you're building a production pipeline around any model here, verify current specs directly with the vendor before committing.