October 2026. Per the roster scan dated October 11 behind this update, the shipping roster holds five lines, and only two of them, Runway Gen-4.5 and LTX 2 Pro, are new since July. Of the six models this article ranked then, the scan still carries Kling 3.0, Veo 3.1, and Sora 2; seedance, Wan 2.6, and SkyReels V4 do not appear in it, which is why the ranking below stays dated instead of claiming a fresh certification. The arrivals column behind the roster runs ten entries deep counting variants, and one of those arrivals, Kling 4.0, has inference vendors publishing launch-day testing guides before Kuaishou has published the model. This update keeps the July ranking as the last fully scored picture of the field, and adds a challengers section shorter than the hype, because the new names have no scorecards yet.
Q4 challengers to watch
Kling 4.0 is reportedly near launch: WaveSpeed, an inference provider in the Kling ecosystem, published a launch-day testing guide on September 27. That is preparation by third parties, not an announcement by Kuaishou. Until the model ships, Kling 4.0 stays a rumor with a date attached, and I don't score rumors. If it arrives on the 4K, multi-shot trajectory Kling 3.0 set, it lands straight in the all-rounder conversation; that's a forecast, not a measurement, re-measured in January.
Behind it, the arrivals column of the October 11 scan:
- Seedance 2.5 and Seedance 2.5 Video Edit. The 2.5 specs (30-second clips, native 4K, up to 50 reference inputs) have been public since the June 23 Volcano Engine FORCE keynote; the dedicated Video Edit variant is the piece that wasn't on the board in July.
- Wan 2.5 / WAN 3.0. The Wan line's circulating version numbers don't reconcile with July, when this article called Wan 2.7 "just shipping." I'm repeating the scan's names, not ranking them.
- MiniMax H3 and H3 Max.
- Gemini Omni Flash / 1.1.
- Flux 3 Video, out of the Flux image lineage.
- HeyGen Avatar 4 and other adjacent releases around the video stack.
The honest caveat, and why this section is a list instead of a ranking: the scan carries names, not scorecards. No verified prices, and no Elo or clip-length ceilings for anything above. I'm not padding it with launch-week leak chatter; the numbers don't exist yet. Every challenger stays unranked until it publishes something dated, and the re-review is booked for January 15, 2027.
Settled: the shipping roster as of October 11 is Kling 3.0 and Kling O3 Pro, Veo 3.1 and 3.1 Fast, Runway Gen-4.5, and LTX 2 Pro. Gone for good: Sora 2 and Sora 2 Pro — the app shut down April 26, 2026, and the API followed on September 24, per OpenAI's own discontinuation notices. Dodged: seedance, Wan 2.6, and SkyReels V4 don't appear in that veteran column. An omission from a third-party digest is not a deprecation notice, so I won't call them retired; I also can't re-certify them past July, which is why the snapshot below keeps its original date.
The mid-2026 snapshot: six names, still standing
The July field: Kling 3.0 (Kuaishou), Veo 3.1 (Google DeepMind), Sora 2 (OpenAI), seedance (ByteDance), Wan 2.6 (Alibaba), and SkyReels V4 (Skywork AI). Video splits along three axes: duration (5s vs 30s), prompt fidelity (does the clip match what you typed), and consistency (does the same face hold across frames and cuts). Almost no model wins all three. The Chinese-trained ones (Kling, seedance, Wan, SkyReels) keep a real edge on Chinese prompts that Western vendors haven't matched. SkyReels V4 took #1 on Artificial Analysis' Text-to-Video (With Audio) board on March 19, beating Kling 3.0, Veo 3.1, Sora 2, Vidu Q3, and Wan 2.6. That's the dated anchor everything below hangs from.
Kling 3.0 (Kuaishou): The All-Rounder
Kling 3.0 is Kuaishou's flagship, released February 5, 2026, and on paper the most capable all-rounder of the year. It generates natively at 4K, runs up to 60fps, extends to 15 seconds, and plans multi-shot sequences instead of one continuous take. "Omni" mode turns text, image, or reference inputs into video. A 10-second standard clip runs about $2.90, a 15-second pro clip with native audio about $8.70, and $0.075-0.168/sec on fal depending on tier. Multi-shot prompts bind characters as @Element references; Kling 3.0 Turbo (June 2026) folded native multi-language audio into per-second pricing; Chinese prompts run closer to intent than Sora or Veo; and motion realism sits near the top of public leaderboards at Elo ~1,248 on Artificial Analysis. The weakness is speed: longer generations take their time, and third-party API throughput can be uneven. It's the pick for longer clips, multi-shot, and image-reference briefs, and if Kling 4.0 ships what the ecosystem expects, it rewrites this section first.
Veo 3.1 (Google DeepMind): The Cinematographer
Veo 3.1 is the one you reach for when the shot needs to look like a movie: shallow depth of field, film grain that makes a clip feel expensive. It outputs at 720p, 1080p, and 4K, max 8 seconds, and ships native synced audio (environment, dialogue, score) without a separate pass. Google's official rate is $0.75/sec on Vertex AI and the Gemini API; third-party endpoints run the same model at $0.09-0.15/sec, where most production traffic actually lands. Camera language ("slow dolly in," "tracking shot") survives the trip from prompt to screen, and per-frame realism on faces and environments is elite. The failure mode is obedience: ask for something mundane and specific, and Veo hands you a beautiful version that drifts from the brief. Chinese isn't its strength, and 8 seconds is the ceiling now that seedance 2.5 goes longer and Sora 2 Pro's 20-second takes retired along with the service.
Sora 2 (OpenAI): The Prompt-Understanding Benchmark, Retired
Sora 2 was OpenAI's video model, and it's still the first name people bring up — which is the awkward part, because it's gone. The web and app experience shut down April 26, 2026; the API followed on September 24, taking Sora 2 and Sora 2 Pro with it. The pricing card, for the record: $0.05/sec at 720p, $0.15/sec for Sora 2 Pro at 720p, up to $0.35/sec at 1080p. Pro's real edge was 20-second single takes with persistent character IDs across calls, the closest thing to "cast the same actor twice" that video generation managed, with audio bundled into every tier. Long, descriptive prompts got interpreted faithfully, and physics and object permanence held up in ordinary scenes. The weakness was cost: at $0.35/sec for 1080p Pro it was the most expensive model here, hard to justify against Wan 2.6 at $0.07/sec for high-volume short-form, and Chinese prompts worked but felt translated.
seedance (ByteDance): The Audio-Video Pioneer
seedance-2.0 is ByteDance's video model, the audio-video pioneer of 2026. It auto-plans shots and camera moves, generates native lip-synced audio in 8 languages, and produces up to 2K cinematic video in about 60 seconds. Pricing is tokens-based: roughly 1 CNY/sec (~$0.14/sec) for pure generation, cheaper for video-edit. Elo 1,269 on Artificial Analysis put it among the board's best, though global API access was China-only through Q2. seedance 2.5 followed on June 23, 2026 at the Volcano Engine FORCE conference with the specs listed above. The sweet spot is 5-15 second clips turned around in 60-120 seconds, Chinese prompts included, at a fraction of Sora or Veo's per-clip cost, which is the whole ballgame at thousands of shorts. It's not the cinematic king; for a 15-second film-festival shot, Veo or Kling will beat it, and resolution caps at 720p/480p on 2.0 (2.5 lifts this to native 4K).
Wan 2.6 (Alibaba Tongyi): The Reliable Budget Option
Wan 2.6 is Alibaba's Tongyi Wanxiang video model, the other major Chinese-trained entry. At $0.07/sec on Atlas Cloud, it's the cheapest AI video generation model available through any major API, and the quality-to-cost ratio is the point: for the price of a single Sora 2 clip you can generate over 20 seconds of Wan 2.6 video. It outputs 720p/1080p, 2-15 seconds, 30fps, with audio sync and multi-shot narrative, native Chinese scene handling, and Alibaba's infrastructure behind it, which means reliable API throughput. The weakness: it doesn't clearly beat Kling 3.0 or SkyReels V4 on any single axis. A strong budget pick, not a category leader. It's the pick for high-volume, budget-sensitive batches, and the reliability play when the leaders are queued or priced out.
SkyReels V4 (Skywork AI): The Leaderboard Leader
Here's the one I almost left out, and shouldn't have. SkyReels V4 from Skywork AI took #1 on Artificial Analysis' Text-to-Video (With Audio) leaderboard on March 19, 2026, beating Kling 3.0, Veo 3.1, Sora 2, Vidu Q3, and Wan 2.6. It runs 1080p at 32 FPS for 15 seconds of cinema-grade audio-video sync, built on a dual-stream MMDiT architecture that solves audio-visual sync at the model level, not as a post-pass, with multi-frame and grid-image references for character consistency. Pricing runs about $0.12/sec ($7.20/min). The weakness: it's the board leader, which means less real-world battle-testing than Kling or seedance, and a leaderboard is one signal that moves. It's the top-of-board default for audio-video prompts, for as long as the board says so.
The snapshot table (July data, re-pinned)
Rankings are my call after months of real use, cross-checked against Artificial Analysis (artificialanalysis.ai); the board standings above carry their March and July dates, and the October 11 scan publishes no fresh order, so this table keeps its July date rather than wearing a fake October one.
| Model | Vendor | Max resolution | Clip length | Chinese-friendly | Realism | Price/sec | Best job today | |-------|--------|----------------|-------------|------------------|---------|-----------|------------------| | SkyReels V4 | Skywork AI | 1080p | 15s | Native | Top-tier (#1 board) | ~$0.12 | Audio-video, top-of-board | | Kling 3.0 | Kuaishou | 4K | 15s | Native | Top-tier | $0.075-0.168 | All-rounder, multi-shot | | Veo 3.1 | Google DeepMind | 4K | 8s | Weak | Top-tier (cinematic) | $0.09-0.75 | Cinematic, hero shots | | Sora 2 / 2 Pro | OpenAI | 1080p | 20s (Pro) | Weak | Top-tier | $0.05-0.35 | English-heavy, long prompts | | seedance-2.0 / 2.5 | ByteDance | 720p-4K | 5-30s | Native | Good (Elo 1,269) | ~$0.14 | Short-form, Chinese, fast | | Wan 2.6 / 2.7 | Alibaba | 1080p | 2-15s | Native | Good | $0.07 | Budget, high-volume |
Caveats: durations and resolutions are typical ranges, not hard caps. "Native" Chinese means first-class Chinese prompt handling; "Weak" means it works but feels translated.
The old problems (hands, consistency, physics)
What still doesn't work, because the demo reels hide it:
Hands. Still. Six fingers, merged fingers, fingers that morph mid-clip. Better than 2024, way better, but if you're watching for it, you'll catch it. Anyone who tells you "hands are solved" is selling you something.
Consistency across cuts. Generate the same character from two prompts and you'll get two different people. Reference-to-video helps (Kling's Omni mode, SkyReels' multi-frame references, Sora 2 Pro's character IDs), but holding a character across a multi-shot sequence is still more art than science, and it's the biggest blocker for "AI filmmaking" being real in 2026.
Physics. Mostly fine for everyday scenes, occasionally horrifying for complex contact: pouring liquid, cloth folding, hands on objects.
None of these are dealbreakers for short-form content (memes, social clips, B-roll). They bite when you need repeatable long-form output. We're not there yet.
Video inside x-rush
None of the six snapshot names, and none of the Q4 challengers, is a badge you'll see inside x-rush, and that is deliberate. Video Studio takes the brief (a 15-second Chinese product clip, a cinematic hero shot, a batch of short-form cuts), Workflow Copilot completes the generation brief, and the credit price is fixed before anything executes. Video generation bills at a flat rate regardless of the engine serving it, and every result is kept with the project. Which engines serve a given video task is an internal decision that tracks the same churn this article describes; you judge the clip, and the scorecard above stays advice for direct API work.
Picking one for direct API work
The last fully scored picture is still July's; October added a roster, not scorecards. The July rule, unchanged: short-form and Chinese prompts go to seedance-2.0 or Wan 2.6, because fast and cheap in Chinese is the job most people actually have. Cinematic hero shots go to Veo 3.1. Longer clips, 4K, and image-reference work go to Kling 3.0, the most capable single pick of the year so far. English-heavy, paragraph-length prompts go to Sora 2 Pro, and you'll pay for it. Audio-video sync goes to SkyReels V4. Budget volume goes to Wan 2.6.
The trap: picking the model with the most viral demo reel, then finding out it can't do the 5-second Chinese clip you actually need. When a challenger ships a real scorecard, weigh it against this table, not against its keynote.
The only stable fact about this field
The 2026 video field moves faster than any other modality, text and image included. Three of July's six names sit in the scan's current roster; the other three it cannot certify, which is why the ranking above keeps its July date. No single model wins all of video, and the challengers get a vote the moment they publish numbers.
The only defensible definition of "best" in a field that reinvents itself every quarter is "best at this job, this month," and you should expect the answer to change under you.
Pin the model to the job, and re-read the board when the challengers ship numbers. Or open Video Studio, describe the shot, and watch the credits come back quoted before the first frame renders.