benchr
Start with the market change, because it resets this roundup. For a year, the assumption was that OpenAI's Sora would define AI video. It won't. OpenAI's Help Center says the Sora app and website retired on April 26, 2026, while the Sora API is scheduled to shut down on September 24, with Sora 2 already marked Legacy. The useful question now is which supported tool fits your production constraints.
Veo 3.1: the native-audio and 4K candidate
Google's Veo 3.1 is the current flagship, and its edge is sound. It generates synchronized native audio, dialogue, effects, and ambience, always on and produced jointly with the picture, so you're not dubbing a silent clip after the fact. Clips run 4, 6, or 8 seconds, and Scene Extension chains them into videos a minute or longer. It does up to 4K at 24fps, takes reference images for character and scene consistency, and supports first-and-last-frame transitions. You can use it in the Gemini app, in Google's Flow filmmaking tool, and through the Gemini API.
Google notes that Veo 3.1 costs the same as Veo 3 did. The Fast tier drops to as little as $0.10 per second and the Lite tier to $0.05, which creates a lower-cost prototyping path before a 4K finish. Its provider-documented feature set belongs on the shortlist when prompt adherence, visual quality, and synchronized audio matter, but those outcomes still need measurement on your prompts. The audio question connects to the wider voice models comparison, while the surrounding Gemini stack is covered in the Gemini evaluation.
The rest of the field
Veo is one serious candidate, not a measured overall winner here. The other tools expose different feature and deployment trade-offs that should determine which ones enter your workload test.
| Tool | Current model | Native audio | Where to use it |
|---|---|---|---|
| Google Veo | Veo 3.1 | Yes, always on | Gemini app, Flow, API |
| Runway | Gen-4.5 | Video-first | Runway app and API |
| Kling | Kling 3.0 | Yes | Kling app and API |
| Luma | Ray3.14 | Video-first | Dream Machine, API |
| Pika | Pika 2.5 | Video-first | pika.art, iOS |
| OpenAI Sora | Being discontinued | — | Retiring through 2026 |
Runway positions Gen-4.5 around visual fidelity, physics, and motion control. Kling 3.0 offers native audio and pushes clip length toward 15 seconds. Luma's Ray3.14 brings native 1080p across its Dream Machine workflows, while Pika 2.5 targets quick consumer clips. Those are vendor features and product positions, not comparable quality scores. Include the candidates that meet your duration, resolution, audio, access, and budget constraints, then judge their outputs blind on the same prompt set.
What none of them can do yet
Be honest with your expectations, because the demos oversell. Every tool here still generates short base clips, from a few seconds up to around 25, that you extend by stitching. Long-form coherence is shaky. On-screen text comes out garbled more often than not. Hands and fine physics still betray the model. And keeping one character consistent across several shots is a real fight, even with reference images. These are clip generators, not film studios, and pretending otherwise is how you waste a budget.
Shortlist Veo 3.1 when synchronized audio, 4K, and Google access are requirements; shortlist Runway Gen-4.5 for motion-control workflows and Kling 3.0 when native audio from a non-Google tool matters. Then use a held-out set of representative prompts, identical duration and aspect-ratio settings, and a blind rubric covering prompt adherence, temporal consistency, motion artifacts, audio sync, latency, and total cost. This page does not name a universal quality winner. Skip Sora for new dependencies because it is being retired, and treat every remaining tool as a source of clips to assemble rather than finished films.
Calculate your cost →·Compare this model →·Find your model →