AI video, after the hype.
Ten models. One verdict.

Sora is the model that lit the fuse. The original Sora landed in December 2024 as a research preview; Sora 2 followed in September 2025 inside a dedicated, social-feeling iOS app — clips up to a minute, native audio dialogue and ambience, and a cameo system that lets the model render a specific person's likeness on demand.

On a good prompt, Sora 2 still produces the most cinematic single takes of any model in this guide. Prompt adherence on hard briefs — moving cameras, multiple subjects, consistent lighting across a shot — beats most competitors. The catch in 2026 is access. OpenAI has been consolidating its product surface, the dedicated apps are no longer the recommended path, and the standalone API has been on a public wind-down trajectory. Treat any "stable Sora pipeline" claim with skepticism until you have written contracts in hand.

Veo 3 was the first frontier model to take native audio seriously, and Veo 3.1 cemented Google's lead there. Dialogue, score, ambience, and Foley arrive in the same file you generated — no post-pass through ElevenLabs or a stock library. 4K output is real, not upscaled, which matters the day a client asks for broadcast deliverables.

The trade-off is the eight-second clip cap. Long-form means stitching clips inside Flow, Google's filmmaking surface — workable for cuts, painful for a single uninterrupted take. Pricing also tilts enterprise: real production work tends to converge on the $250/mo Ultra tier or a Vertex AI contract. Both are fine if you are an agency; both feel like a wall if you are a solo creator.

Runway has been doing this longer than anyone, and Gen-4.5 keeps the company's reputation for character and world consistency intact. Build a reference image, drop it into a scene, and you can carry that face, that costume, that lighting across multiple shots without the model forgetting. That alone is why agencies running episodic or branded character work still default to Runway.

Aleph is the second weapon: a real video-to-video model that takes existing footage and re-renders it in a new style, with a new subject, under new lighting — without the temporal jitter that killed early V2V attempts. Combined with Act-One and Act-Two for performance capture, Runway has assembled the most production-mature creative surface in the market. The catch is credit economics: iteration burns currency fast.

Kling is the model the rest of the industry quietly downloads to benchmark against. Two-minute generations at 1080p / 30fps are uncommon — most competitors cap clips at 10 seconds, then ask you to stitch. Kling renders the long take. Physical realism is also disproportionately strong: water, hair, fabric, cloth simulation, the small physics the eye flags first.

The friction is operational. Kling's English-language UX has improved but historically required Chinese-side onboarding. There is no native audio path — every Kling clip needs sound mixed in after the fact. Content policy is real but opaque, and shifts without much warning. If your compliance team will not greenlight a Chinese pipeline, Kling is off the table regardless of how good the renders are.

Pika took a different bet from the frontier labs: instead of chasing benchmark wins, it built the lowest possible friction from prompt to first usable clip. Pikaffects — one-tap effects like "explode this", "melt this", "inflate this" — let non-technical users get something fun in 30 seconds. Pikaformance, the current generation, adds convincing lip-sync to arbitrary audio.

That is also the ceiling. Pika is not the tool you reach for when the brief is broadcast or long-form cinematic. Prompt adherence is good not great compared to Veo or Sora; output ceiling sits below open-source LTX-2 on hero shots. But for social-first content, talking-head reactions, and meme-velocity output, no competitor matches Pika's time-to-first-good-clip.

Luma sells a system, not a model. Ray 2 is the core video engine — fluid, physically plausible motion that handles momentum and inertia better than most peers. But the actual product is Luma Agents, an orchestration layer that runs image, video, audio, and copy generators against a single brief, holding consistency across them.

The pricing ladder — $30 / $90 / $300 per month — makes procurement a one-page conversation, which is rare in this category. Named agency customers (Publicis, Serviceplan, Dentsu) suggest the team has built the contractual surface that agencies actually need. The trade-off: if you only need single text-to-video and want the absolute lowest cost-per-clip, Luma is not the cheapest seat.

Hailuo's bet is verticalisation. Instead of one prompt box, the platform ships a dozen pre-built tools — PetPal, BabyForm, ASMR Generator, Style Switch, CineScope, Ads Generator. Each one wraps the model in a use-case-specific UI that strips out prompt-craft entirely. For consumer-creator audiences this is enormous; for prompt-fluent power users it is incidental.

The underlying model is also genuinely strong on physical realism, particularly faces and bodies. Hailuo Agent automates multi-tool workflows — generate the still, animate it, swap the style, ship. The English documentation lags the Chinese-language version, and enterprise contracting outside China is harder than with Western vendors. As a creator product, that is fine. As an agency procurement decision, it is friction.

Seedance 2.0 was the first big release of 2026. It is also the most controversial: shortly after launch, public reporting flagged copyright-infringement concerns on training data, which is the kind of disclosure that turns enterprise procurement teams cold immediately. For consumer-creator use the question is fuzzier; for any commercial work where IP indemnification matters, it is decisive.

On the model side, Seedance is unusually good at human motion, dance especially — the training signal is obvious if you watch a few outputs. Five to fifteen seconds is the operating range. Turnaround is quick versus frontier peers. If you live inside the ByteDance ecosystem already (TikTok-adjacent pipelines, CapCut), the integration story is real. Outside it, the legal cloud is the deciding factor.

LTX-2 is the open-source moment of 2025 — open weights, 60-second native generations, and audio built into the model rather than bolted on after. If you have GPUs and an engineering team, you can stand up an in-house pipeline today that escapes per-clip pricing entirely. That is not just a financial argument; it is a sovereignty argument for anyone whose data cannot leave the building.

The trade-offs are honest. Self-hosting requires real GPU infrastructure to be production-grade; "runs on a consumer card" and "runs at production throughput" are different sentences. The hosted LTX Studio offers a managed path at a cost below Runway's, but with fewer creative-control surfaces. Output ceiling on hero shots still sits below Veo 3.1 or Sora 2. For most enterprise teams, LTX-2 is the right second pipeline, not the only pipeline.

Wan is the other open-source story worth taking seriously. Alibaba has been releasing checkpoints on Hugging Face and ModelScope with a cadence and quality that surprised the open community. Multilingual prompt understanding is a quiet strength — Mandarin briefs translate to clean visual intent in a way most Western-trained models still fumble.

Quality on image-to-video is the headline. On a single still, Wan produces motion that looks frontier-adjacent — the first time the open ecosystem has been able to make that claim with a straight face. Capping at ten seconds is a real limit, and going from research checkpoint to production pipeline is the same engineering project as with any open model. If your team is Mandarin-fluent or self-host-capable, Wan belongs in your stack.