What's New in AI Video Models: 2026 Mid-Year Roundup (Sora 2, VEO 3.1, Kling 3)

The AI video field has shifted twice since January. If you read our best AI video generation models 2026 post in April, three of those leaderboard entries already have new versions, two new players have appeared, and the entire pricing landscape has rebalanced around audio-native generation.

Here's the honest mid-year state of things — what shipped, what regressed, and what to actually use as of May 2026.

What changed since the last roundup

Four real shifts to internalize before you re-tool:

1. Audio-native is now the table stakes. Sora 2, VEO 3.1, Kling 3.0 and PixVerse V6 all generate synchronized dialogue, foley and music in a single pass. If your current model still requires a separate ElevenLabs round-trip for voice, you're paying double and losing lip-sync fidelity.
2. The Chinese stack caught up — and in some cases passed. Kling 3.0 (Kuaishou) and Hailuo 2.3 (MiniMax) both rank ahead of VEO 3 on Artificial Analysis's video Elo, and Wan 2.7 is the strongest open-source release ever shipped.
3. Sora 2 went paid-only. Free Sora generation ended on January 10, 2026 (Apiyi). If you were running unpaid pipelines, they broke months ago.
4. VEO 4 is still vapor. Despite a flood of "Veo 4 release date" SEO content, Google DeepMind has not announced an official Veo 4 model page or API ID as of April 2026 (Quest Studio). Plan around 3.1.

Sora 2 (OpenAI — September 2025, still flagship)

Sora 2 is the model that forced everyone else to ship audio. Officially announced on September 30, 2025, it pairs synchronized dialogue, sound effects and ambient noise with sharper physics than Sora 1 — basketballs that miss the rim now rebound instead of teleporting through it (OpenAI).

The hard numbers:

• Duration: 10–25 seconds per clip
• Resolution: 720p (Sora 2) or up to 1024p via API / 1080p via subscription (Sora 2 Pro)
• API pricing: $0.10/sec at 720p, $0.30/sec for Pro 720p, $0.50/sec for Pro 1024p (WaveSpeed)
• Subscription: ChatGPT Plus ($20/mo, 480p unlimited) or ChatGPT Pro ($200/mo, 10,000 credits up to 1080p)
• Disney partnership: licensed character generation now permitted in custom scenarios

Where it wins: Physical realism, comedy timing, "self-insertion cameos" via the Sora iOS app, and any prompt where momentum and physics matter (sports, slapstick, stunts). The system card emphasizes steerability — Sora 2 follows multi-shot direction better than any Western competitor.

Where it loses: Cost. At $0.50/sec a 25-second 1024p clip is $12.50. A Hailuo 02 clip at the same length is roughly $0.70. Also: 25 seconds is the hard cap. There is no scene-extension feature.

VEO 3.1 (Google DeepMind — January 13, 2026)

VEO 3.1 was the January launch that closed the dialogue gap. The headline feature is Ingredients to Video: upload up to three reference images (a character, a location, a prop) plus a prompt, and VEO weaves them into a coherent clip while keeping identities locked across cuts (Google blog).

What 3.1 added on top of VEO 3:

• Native 9:16 vertical for Shorts and TikTok (no center-crop hacks)
• 4K upscaling from a 1080p base
• Scene Extension for narratives over 60 seconds
• Synchronized multi-person dialogue with timing-accurate sound effects
• Available in Gemini app, Flow, Vertex AI, Gemini API, Google Vids, and YouTube Shorts (CineD)

VEO 3.1 is the right pick when you need a recurring character to deliver lines with believable lip-sync. We covered the practical side in our VEO 3.1 complete guide. For the head-to-head with Sora, see Sora 2 vs VEO 3.1 deep capability comparison.

Kling 3.0 / O3 (Kuaishou — February 2026)

Kling 3.0 (also surfaced as "Kling O3") is the first unified multimodal video model — video, audio and image generation share a single architecture instead of being chained together. That means native lip-sync, multi-shot storyboarding, and element consistency cooperate by design rather than as patched-on tools (MindStudio).

The intermediate Kling 2.6 release (December 3, 2025) was the first to ship synchronized audio for the family — 1080p at 48 FPS, 10-second max clip, plus the Elements feature for combining up to four reference images for character consistency (Artlist).

Why it matters: Kling 3.0 currently ranks at or above VEO 3 on community benchmarks and is the model to beat for cinematic motion plus audio in a single shot. The 10-second cap is the catch — you'll need to chain via first/last-frame control to build longer scenes.

Hailuo 2.3 (MiniMax — 2026 update)

Hailuo 02 dropped in June 2025 and was instantly the price-performance winner. The 2.3 update kept the same $0.28/clip pricing while improving dynamic expression and stability (Hailuo 02).

• 1080p native
• 6 or 10-second clips at 24–30 FPS
• Currently #2 on Artificial Analysis's video leaderboard, ahead of VEO 3 (MiniMax)
• "Extreme physics" branding is real — water, fabric and crowd scenes hold up

If your workflow is volume-driven (UGC ads, product B-roll, niche YouTube), Hailuo is still the cost-per-clip winner by a wide margin. We've baked it into our AI B-roll generator for exactly this reason.

Runway Gen-4.5 (March 2026)

Runway's Gen-4 launched March 2025 with the world-consistency pitch — characters, locations and props that survive cuts. Gen-4.5 ships in 2026 with state-of-the-art motion quality, prompt adherence, and visual fidelity (Runway).

The unique selling point is still production-grade controls: real camera-motion language, light-direction control, and the closest thing to a virtual cinematography rig in any of these tools. If you're cutting AI footage into traditional editorial, Runway slots in cleanest because the controls map to a DP's vocabulary.

PixVerse V6 (March 30, 2026)

PixVerse V6 is the dark-horse release of the spring. Twenty-plus cinematic lens controls (focal length, aperture, depth of field, lens distortion, chromatic aberration, vignette), native multi-shot generation with synchronized audio, and — uniquely — a CLI that plugs into Claude Code, Codex and Cursor for agent-driven video pipelines (PixVerse).

V6 marks PixVerse's transition from a template-heavy app to a genuine production platform. If you're scripting batch generation, it's the first model where the developer ergonomics match the visual quality.

Wan 2.7 (Alibaba — April 2026, open source)

Open-source video had a quiet 2025; April 2026 changed that. Alibaba shipped the full Wan 2.7 suite — text-to-video, image-to-video, reference-to-video with voice cloning, and instruction-based video editing — all under Apache 2.0 (Cliprise).

Wan 2.5 already supported native synchronized audio (voice + SFX + lip-sync in one step) at 480p/720p/1080p with 10-second clips. 2.7 adds first-frame control and 15-second clips. Alibaba has confirmed Wan 3.0 will target 4K, 30-second generation, and 60B parameters under the same open license, mid-2026.

For self-hosters and anyone who needs commercially-permissive weights, this is the only credible option at this quality tier.

Two more shifts worth tracking

Beyond the model-by-model news, two structural changes in the field deserve their own attention.

The Artificial Analysis leaderboard has become the de-facto reference. Throughout 2025 it was easy to dismiss benchmarks as either vendor-curated or hobbyist. As of mid-2026, Artificial Analysis's video Elo, with thousands of pairwise human votes per model pair, is the best public signal — and it currently puts Hailuo 02 at #2, ahead of VEO 3, with Kling and Sora 2 trading the top spot depending on the prompt category. If a vendor cites a different benchmark, ask why.

Cost-per-second is collapsing — but not uniformly. Hailuo 2.3 at $0.28/clip, Wan 2.7 self-hosted at near-zero marginal cost, and PixVerse V6 batch pricing have all pushed the floor down. Sora 2 Pro at $0.50/sec for 1024p is now the high-end outlier, not the norm. If you're building a content business, the unit economics in May 2026 are dramatically better than in January — but only if you're willing to model-shop per shot.

Open-source caught the closed models. Wan 2.5 with native synchronized audio, then 2.7 with first-frame control and 15-second clips, means open weights are now within striking distance of VEO 3.0 quality. For self-hosters, hobbyists, and anyone with compliance constraints around closed APIs, this is the first time in two years you don't have to compromise on quality to stay open.

Practical takeaway: what to use when

Use case	First pick	Why
Dialogue-heavy character scene	VEO 3.1 (Ingredients to Video)	Best identity lock + lip-sync
Physical-action / sports / comedy	Sora 2 Pro	Physics realism, momentum
Cinematic short film, single shot	Kling 3.0	Unified multimodal, 1080p/48fps
High-volume UGC ads	Hailuo 2.3	$0.28/clip, fast
Editorial-friendly camera language	Runway Gen-4.5	DP vocabulary, motion controls
Multi-shot with audio in one prompt	PixVerse V6	Native multi-shot, CLI
Self-hosted / open weights	Wan 2.7	Apache 2.0, native audio

The honest meta-strategy: don't pick one. The reason we built Versely's AI video generator as a multi-model router is that VEO is wrong for slapstick, Sora is wrong for budget B-roll, and Hailuo is wrong for branded character work. You want a single billing surface that lets you hop models per shot.

FAQ

Is Sora 2 still the best AI video model in May 2026?

For physical realism and steerability, yes. For everything else — dialogue, vertical, character consistency, cost — there's a better answer. Kling 3.0 and VEO 3.1 trade wins on Artificial Analysis benchmarks depending on the prompt category.

Has Google released VEO 4 yet?

No. As of early May 2026, Google has not officially announced VEO 4, published a model page, or shipped an API ID. The "Veo 4 April 2026 release" content circulating on lower-quality SEO sites is speculation. VEO 3.1 remains current (Quest Studio).

Can I still use Sora 2 for free?

No. As of January 10, 2026, Sora image and video generation is gated to ChatGPT Plus ($20/mo) and Pro ($200/mo) subscribers. Free-tier access ended.

What's the cheapest model that still produces broadcast-acceptable video?

Hailuo 2.3 at roughly $0.28 per clip. Wan 2.5/2.7 if you can self-host on an H100 or rent it. PixVerse V5.6 sits in the middle of the price-quality curve.

Does any model do longer than 25 seconds in a single generation?

PixVerse V6 multi-shot and VEO 3.1 Scene Extension both push past 60 seconds in a single workflow, but they're stitching clips under the hood. True monolithic 60-second generation is not yet shipped at consumer-grade quality.

Workflow patterns we're seeing in May 2026

Three patterns have emerged across the creator teams we work with this quarter:

The two-model A/B for every shot. Send the same prompt to two models (commonly VEO 3.1 and Kling 3.0, or Sora 2 and Hailuo 2.3), pick the winner. Cost is roughly 2x but iteration time drops because you don't have to re-prompt when the first generation misses. Worth it for hero shots, overkill for B-roll.

Image-first, then image-to-video. Generate the keyframe in Midjourney V8.1 or FLUX.2 first, then pass it as a reference into VEO 3.1 Ingredients or Kling Elements. This gives you art-direction control on the visual aesthetic before committing video credits. We covered this pattern in AI image to video vs text to video: which to use.

Audio in-prompt, no post-render. Now that Sora 2, VEO 3.1, Kling 3.0 and PixVerse V6 all generate audio in-pass, the old workflow (generate silent video → ElevenLabs voiceover → align in Premiere) is being abandoned. The in-pass audio isn't always better than ElevenLabs v3, but the time savings dominate for short-form social content where the cost of a 10% lower voice quality is outweighed by 3x faster turnaround.

The next move

Mid-year is when most creators rebuild their model stack. If you've been hand-rolling between four browser tabs to get a finished video, that's the symptom — the fix is a multi-model pipeline behind a single interface. Try Versely's AI video generator or build a full short with the AI movie maker and decide for yourself which model wins each shot.

For the deeper Sora-vs-VEO breakdown, read Sora 2 vs VEO 3.1 deep capability comparison. For an end-to-end production playbook, see AI content creation 2026 complete playbook.