What is Wan 2.6?

Wan 2.6 is Alibaba's (Tongyi / Wan team) multimodal video generation model, released December 2025. It turns a text prompt — or images, a reference video, and audio — into a clip up to 15 seconds long at up to 1080p (24fps), with audio generated in the same pass: dialogue with lip-sync, sound effects, and music.

Over Wan 2.5 it adds three things that matter for real production. Multi-shot narrative planning: one prompt can lay out several cuts (wide → close-up → reaction) and the model sequences them, instead of you generating and stitching separate clips. Reference-based generation: supply reference images or a reference video and the model holds a character's identity and look across shots. Higher motion fidelity: smoother, more stable motion across the longer 15s runtime. It accepts text, image, reference-video, and audio inputs in one workflow.

Wan 2.6 is an API-only model — weights are not publicly downloadable. Wan 2.1 and 2.2 are the open-weight (Apache-2.0) versions; 2.5 and 2.6 are API-only. On GPTProto you call it with the model string wan-2.6, billed per run by resolution and duration.

How multi-shot and synced audio work

Two capabilities define Wan 2.6, and both change how you prompt.

Synced audio in one pass. The generation step produces frames and a matching audio track together, so speech lands on the right lip movements and ambient sound and music sit under the cut without separate editing. Describe the sound in your prompt with a Sound: cue, or pass your own track via the audio parameter and the model syncs motion to it.

Multi-shot in one run. Rather than rendering one continuous shot, Wan 2.6 can sequence several beats inside a single clip. Segment the prompt by time ([0-5s] … [5-10s] … [10-15s] …) and the model treats each as a shot, carrying characters and setting across the cuts. The shot_type parameter controls whether you get a single continuous shot or a multi-shot composition. For coherence, 8–12s clips tend to be the most stable; 15s is available when you need the length.

Choosing inside the Wan family

GPTProto carries the Wan models on one balance. Pick by the job:

• Wan 2.6 — text-to-video (this page): 1080p, up to 15s, multi-shot, native audio. Use it for longer or multi-cut narratives from a prompt.
• Wan 2.6 — image-to-video / reference-to-video: animate a still, or guide a scene with a reference clip / up to several reference images for identity. 
• Wan 2.5 (wan-2.5): 1080p, up to 10s, native audio, single-shot — the cheaper daily driver, from $0.225/run.
• Wan 2.2 (wan-2.2-plus): silent, 720p, ~5s — but open-weight, the pick when you must self-host.

Competitors list these models without a selection map. The short version: 2.6 for length + multi-shot + identity, 2.5 for cheaper single-shot clips with audio, 2.2 for open weights.

Wan 2.6 vs Wan 2.5 vs Wan 2.2

Honest take: Wan 2.6 costs more per run, but it's the one to use when you need 15s, multi-shot scenes, or character consistency across cuts. Wan 2.5 is the cheaper daily driver for single-shot clips up to 10s with audio. Wan 2.2 is the pick only if you need open weights to self-host.

Related: Wan 2.5 → · Wan 2.2 → · Sora 2 →

What you can build with the Wan 2.6 API

Multi-shot short films and ads — One prompt lays out several shots with consistent characters and audio across cuts — a 15s narrative ad without manual stitching. A 20-variant test at 720p/5s ≈ $9.00.

Localized video at scale — Multilingual prompts and lip-synced speech turn one storyboard into several languages without re-shooting. Strong Chinese support.

Any aspect ratio, same price — Square (960×960, 1440×1440), portrait (720×1280, 1080×1920), and landscape are priced identically within a tier, so format is a free choice per platform.

Image / reference-to-video — Animate a still or guide a scene with a reference clip on the sibling endpoints. 

Wan 2.6 prompt recipes

Wan 2.6 generates audio and can plan multiple shots, so prompts work best when they describe shots, motion, and sound together. The platform's own example segments the prompt by time — use that for multi-shot. Structure each beat: shot + subject + action + camera + lighting + Sound: cue.

1. Multi-shot narrative (15s + shot planning)

[0-5s] Wide shot: a lone hiker reaches a misty mountain ridge at dawn, slow push-in.
[5-10s] Medium shot: she lifts her camera and smiles. Sound: wind, distant birds.
[10-15s] Close-up: the sun breaks over the peaks, light fills the frame. Sound: a
soft swell of ambient music, no dialogue.

2. Dialogue + lip-sync (single shot)

Medium close-up of a barista behind a wooden counter, warm morning light. She looks
to camera and says, "Your usual? Coming right up." Steam rises from the cup. Sound:
spoken line, espresso machine hiss, quiet cafe ambience.

3. Product/promo, square frame

Square frame. A sneaker rotates slowly on a matte pedestal, studio light sweeps
across it, subtle reflections. Sound: a low synth pulse, soft whoosh on the sweep,
no speech.

Working tips: for multi-shot, label each beat with its time range and keep dialogue short enough for the beat; set shot_type for single vs multi-shot; 8–12s is the sweet spot for coherence, 15s when you need length; use negative_prompt to exclude artifacts (e.g. "no text, no watermark"); prompt_extend: true lets the model expand a short prompt — turn it off for exact control; fix seed to reproduce a result.