Generative Video Drops! Kling 2.6, o1, & NEW Models!

Kling 2.6 arrives with native audio generation, and early tests suggest its sound quality is competitive with top-tier rivals—while still showing the familiar “not-quite-cinema” grain and occasional audio quirks. The biggest practical shift is that audio now comes from the model itself rather than being bolted on afterward, letting dialogue, music cues, and scene-appropriate effects land more coherently inside the same generated clip. In side-by-side impressions, Kling 2.6’s dialogue delivery is described as clearer than LTX-2 and often on par with Sora 2 or VEO3 for simple speech, with music and environmental sound effects (footsteps, birds, cinematic scoring) that match the visuals closely enough to feel like a movie scene.

The clip set also highlights where the technology still struggles. Lip-sync can wobble, some audio carries a slight uncanny texture, and certain effects show “AI-ness” through graininess or resonance. Visual artifacts show up too—teeth and sword details can morph over time, and even when the model nails the overall concept (lighthouse shots, lava-dragon physics, screaming and roars, Paris romance beats), fidelity isn’t perfectly stable across longer sequences. Still, the model’s ability to hold a 10-second shot with consistent action and to generate coherent sound effects repeatedly is treated as a meaningful step forward.

Pricing and workflow details add another layer: Kling 2.6 is accessed through Fall AI using a credit system and an API, with audio generation priced at about 14 cents per second (roughly $1.40 for a 10-second clip). The transcript contrasts that with Sora 2 at 10 cents per second, while also noting Kling 2.6’s faster generation speed in the tested scenarios.

Beyond Kling 2.6, the discussion widens to the broader race in generative video. Apple’s new Star Flow V experiment is framed as a non-diffusion approach using normalizing flows, with a key differentiator: reversibility. That means text-to-video and video-to-text are both possible, potentially enabling more interactive or real-time applications—though current examples are said to lack the fidelity and real-world understanding of the leading systems.

Kling 01 then takes center stage as a “Nano Banana style” character and scene reconstruction model for video. Demonstrations emphasize preserving backgrounds and continuity while swapping characters, including replacing a dancer with a new person, swapping cars in a James Bond–like scene while keeping surrounding actors intact, and running more complex edits like continuous handheld shots with consistent footwear and camera movement. Comparisons against Runway Aleph suggest Kling 01 can be stronger at prompt adherence in some cases (e.g., medieval armor details, snow accuracy), while Runway may preserve certain background elements (like a kitten) better. The transcript also flags upcoming competition: Runway 4.5 is teased as a video generator with confirmed audio support.

Overall, the throughline is clear: native audio, reversible generation methods, and stronger continuity/editing are converging—turning AI video from “cool clip generator” into a tool that can support more production-like workflows, even if artifacts and occasional incoherence remain.