Updated 2026-08-15 · Checked daily, updated at least weekly
Best AI video models right now
The flagship video generators, ranked by what they're actually best at — with real outputs, blind-vote arena scores and side-by-side comparisons. Rumors labeled, prices sourced, nothing quoted from memory.
Editor's pick
Best all-around today — but blind-vote arenas are a real fight: Seedance 2.0 and Kling 3.0 trade the top spot depending on the test.
Best video model for each job
Best overall
The most steerable flagship: cinema-grade realism, accurate physics and 48kHz native audio.
Native audio · ~4K · premium API
Best long single take
Released Jul 31: 30-second generations that extend into minutes, and 50 reference assets in one request. No API yet — Jimeng and Doubao Pro only.
30s + extension · 50 refs · no API yet
Best for realistic humans
Best-in-class faces and motion with synchronized dialogue, not just SFX. Four entries in the arena top 10.
Multi-shot · lip-sync · 4K
Best for narrative
Long, coherent narratives with synced dialogue. Caveat: the API retires Sep 24, 2026 — plan a migration.
Narrative · ChatGPT Pro · retiring
Best for camera control
Pan/tilt/zoom precision and generative editing that filmmakers rely on.
Camera moves · generative editing
Best value
Native 2K with stereo audio from $0.14/s, and text, image, video and audio read as one context. Launched Jul 31, so quality is unproven by any independent board — and fal doubled its rate on day one, so shop the provider.
2K · native stereo · omni-reference
Best for social short-form
Fast, dynamic, social-ready clips with multi-shot and native audio.
Fast · social · multi-shot
Text-to-video Arena (no audio)
Source: Artificial Analysis · as of Jul 2026
Elo scores are volatile and some leaders (HappyHorse, Seedance 2.0) have limited public access. This is the no-audio board; Gemini Omni Flash also tops the separate with-audio track. Neither FLUX 3 (gated early access since Jul 23) nor MiniMax H3 (launched Jul 31) has appeared here yet, so launch-day ranking claims for both are unverified.
What the newest models actually produce
Official launch clips from the three models that shipped in nine days — sound on.
The longest single generation any flagship offers, unedited and with sound.
From MiniMax's own API docs. The only one of the three you can buy today.
Black Forest Labs' longest published demo; the model itself is still application-gated.
Official launch clips — each vendor's own demo, sourced on its model page below · mirrored and re-encoded here, sound intact.
Prompt and output, paired
Hover to play — every clip ships with the exact prompt that made it (Seedance 2.0, from ByteDance's prompt guide).
Product ad: golden horse pendant
Reference the running form of the horse from Video 1, generate a golden horse galloping on a grassland, then freeze-frame its magnificent running pose, transforming into a horse-shaped gold pendant.
Consistent character, new scene
Reference the woman's appearance from Image 1, Image 2, and Image 3, generate a scene of her eating cake at a coffee shop.
Choreographed fight scene
Reference the character actions and camera language from Video 1, generate a fight scene between Image 2 and Image 1. Image 2 is the character on the left, Image 1 is the character on the right. With intense background music.
Golden particle effects transfer
Reference the golden particle effects from Video 1, have the character from Image 2 play a flute while surrounded by the same particle effects.
FPV dive concept video
Reference the camera movement from Video 1 to create a concept video for a tech park. Use the high-rise building from Image 1 as the visual center, with the same first-person diving perspective, highlighting the tech aesthetic of the park.
Scene-to-scene particle transition
Video 1, at the moment the leaf touches the ground, golden particle effects burst out, a gust of wind blows, then connect to Video 2.
Demo outputs and prompts from ByteDance's official Seedance 2.0 prompt guide. Hover to play. Our own same-prompt tests replace these at each model launch.
Same prompt, every model
Head-to-head breakdowns on the specs that decide a shot — not a beauty contest.
Seedance 2.5 vs H3 vs FLUX 3
Three flagships in nine days. What's verified, what's being repeated, and the launch-week claims that fail a check.
Seedance 2.5 vs 2.0
What actually changes: duration, resolution, references, pricing.
Seedance 2.5 vs Kling 3.0
Long single takes vs multi-shot storyboards — which fits your work.
Seedance 2.5 vs Veo 3.1
China's efficiency play against Google's audio-first flagship.
Video tools & guides
Prompt libraries
Copy-ready prompts for Veo 3 and Seedance — JSON and prose.
Shot Planner
Turn a brief into a shot-by-shot prompt on the official formula.
Cost calculator
Source-linked per-second pricing across providers.
Release calendar
Every confirmed drop and shutdown date — subscribe via .ics/RSS.
Get video-model drop alerts
One email when a major video model ships or gets a confirmed date. No spam, unsubscribe anytime.
Full release tracker
Upcoming & rumored
One network for image, video with native audio (up to 20s) and robot action prediction. Video is open to approved applicants only; there is no public API, no published pricing and no open weights yet.
MiniMax says it will publish H3's weights “within the next few days, subject to the relevant laws and regulations”, and designed the model for compatibility with several domestic Chinese accelerators. No repo has appeared on the MiniMaxAI Hugging Face org yet. Would be the strongest openly licensed video model to date.
ByteDance says the Volcano Ark API is coming shortly after the Jul 31 consumer launch. Until it lands there is no per-second price, no third-party endpoint and no access outside Jimeng and Doubao Pro. The three insider-sourced dates before launch (Jul 6, 13, 16) all failed, so treat any new date without an official post the same way.
Two-stage wind-down of Sora 2: the consumer app and sora.com already went offline on Apr 26, 2026; the developer API (sora-2, sora-2-pro) is decommissioned Sep 24, 2026. Teams on Sora endpoints need a migration path before this date.
Not announced. Estimate based on Kuaishou's historical 4–6 month release cadence since Kling 3.0 (Feb 2026).
Confirmed but undated: the open-weight release of the full multimodal backbone — the first openly licensed model generating video, audio and images from one set of weights. No size or license terms announced; FLUX.1/FLUX.2 Dev tiers shipped non-commercial.
Not announced. Google I/O 2026 shipped Gemini-family updates, not Veo 4. Rumors point to native 4K and longer clips.
No Sora 3 has been announced. After shutting down Sora 2, OpenAI said it is repurposing Sora as a world-model research effort — there is no successor product or date on the roadmap.
Shipped
Hailuo's third generation, launched as a general-purpose omni model: one network reads text, images, video and audio as a single context and returns 2K video with native stereo sound, 5–15s. Live the day it was announced — in the Hailuo app, MiniMax's own API and third-party resellers. Open weights promised within days.
Released Jul 31, 2026 after a Jun 23 preview. 30-second single generations with multi-round extension into minutes, up to 30 images + 10 videos + 10 audio as references in one request, and timestamp-addressed editing (green screen, viewpoint, camera move). Live in Jimeng AI and Doubao Pro; the Volcano Ark API is “coming soon” and no 2.5 model ID exists in the ModelArk catalog yet. No international access, no pricing.
Launched Feb 12, 2026 with Doubao 2.0. Multi-modal input (up to 9 images, 3 videos, 3 audio files), native audio, 480p–4K output. ByteDance paused its global rollout in mid-March 2026 after cease-and-desist letters from Disney, Paramount, Warner Bros. and the MPA over copyrighted characters in outputs.
Multi-shot storyboard generation — one prompt produces a sequence of coherent shots.
Google's current flagship video model. Strong physics and native audio; premium API pricing.