Wan 2.6 API and Pricing, Explained Through a Real Pipeline
Wan 2.6 is the Alibaba video model before Wan 2.7 and Wan 3.0, and on Treza it is the one with the simplest rate card in the family: 8 cents a second at 720p from a prompt, 10 from a still, 12 and 15 at 1080p, with sound generated alongside the picture. Here is the rate card, what our test render billed and returned, how it compares with its two successors, and how to reach it as a versioned API endpoint.
Wan 2.6 is Alibaba's video model from before Wan 2.7 and Wan 3.0. It renders a 5 or 10 second clip in 16:9 or 9:16 from a text prompt, or from a still you hand it as the first frame, and it generates ambience and effects alongside the picture. OpenRouter prices it by resolution and by input: 8 cents per second at 720p from a prompt, 10 cents from a still, and 12 and 15 cents at 1080p. It has had its own page on Treza since October 14, 2026, it sits in the Video Generation node's picker beside Wan 2.7 and Wan 3.0, and you can ask for it by name in chat.
This post covers what Wan 2.6 costs at each setting, what our test render billed and came back as, how it compares with the two newer Wan models, what the endpoint accepts and leaves out, and how to call it as a published API endpoint. Every parameter and price below is read off Treza's live video model catalog as of October 2026, and the behavior described was checked with a real render.
What you get before you talk about price
| Capability | Wan 2.6 |
|---|---|
| Clip length | 5 or 10 seconds |
| Resolution | 720p or 1080p |
| Aspect ratios | 16:9, 9:16 |
| Audio | Generated with the clip; can be turned off per generation |
| Keyframes | First frame only |
| Reference images | No, not on this endpoint |
Two lengths, two shapes, two resolutions, one switch for sound. That is the whole spec sheet, and it covers wide for YouTube and the web and vertical for Shorts, Reels and TikTok. Capability detail lives on the Wan 2.6 model page.
How Wan 2.6 bills
Per second of output, at a rate set by the resolution and by whether the clip starts from a still:
| Input | 720p, per second | 1080p, per second |
|---|---|---|
| Prompt | $0.08 | $0.12 |
| Prompt plus a first frame | $0.10 | $0.15 |
So a 5 second clip from a prompt at 720p is 40 cents of provider cost, and the same clip at 1080p is 60. Starting from a still adds a quarter to the rate: 50 cents for a 5 second 720p clip, 75 cents at 1080p. The most expensive clip the model can render, 10 seconds at 1080p from a first frame, is $1.50.
The audio switch does not change the rate. Sound is part of the listed price, so turning it off saves nothing; it only spares you a track you would otherwise have to duck under a narrator.
For scale, HeyGen Video starts at 2 cents a second at 480p, Veo 3.1 Lite at 3 cents for a silent clip at 720p, Wan 3.0 and Grok Imagine at 5 cents at 480p, and Wan 2.7 at a flat 10 cents at either resolution. Wan 2.6 at 720p from a prompt sits just under its successor, and it is the cheapest Wan rate that includes sound at 720p.
What our test render billed, and what it came back as
We rendered one 5 second vertical clip from a prompt at 720p on October 14, 2026, with audio on: a barista pouring steamed milk into a latte in a sunlit cafe, slow push-in, natural ambient sound. Four things are worth knowing from it.
The bill matched the arithmetic exactly. Five seconds at 8 cents is 40 cents, and that is what the provider charged. No minimum length beyond the 5 seconds you asked for, no rounding up.
The frame was what the setting promised: 720 by 1280 pixels at 30 frames per second, running a hair over 5 seconds. That is worth saying because another model we tested the same week returned a smaller frame than its catalog listed.
It carried sound. The file had an AAC audio track with the room's ambience in it, at a level well below normal listening volume, quiet enough that a voiceover laid over it does not fight for space.
And it was quick for a model with audio. The clip took 33 seconds to come back, against 71 seconds for a silent 6 second clip on Hailuo 2.3 in the same batch of tests.
We have not rendered 1080p, so we cannot say what frame size it returns beyond the listed resolution.
Wan 2.6 vs Wan 2.7 vs Wan 3.0
Same vendor, three models in one picker, and the newer two each buy something specific.
| Wan 2.6 | Wan 2.7 | Wan 3.0 | |
|---|---|---|---|
| Clip length | 5 or 10 seconds | 2 to 10 seconds, exact | 2 to 30 seconds, exact |
| Aspect ratios | 16:9, 9:16 | 5, adds 1:1, 4:3, 3:4 | 5, adds 1:1, 4:3, 3:4 |
| Resolution | 720p, 1080p | 720p, 1080p | 480p, 720p, 1080p |
| Keyframes | First frame | First and last frame | First and last frame |
| Reference images | No | Yes, including real people | No |
| Audio | Generated, can be turned off | Generated, can be turned off | Generated, can be turned off |
| Price per second | $0.08 to $0.15 by resolution and input | $0.10 flat | $0.05, $0.10, $0.20 by resolution |
Wan 2.7 is the people model. It takes reference photos, including recognizable real people, pins a shot at both ends with a first and a last frame, and renders 2, 3 and 4 second inserts that Wan 2.6 cannot. The Wan 2.7 model page covers it, and the Wan 2.7 vs Seedance 2.5 comparison puts it against our default model.
Wan 3.0 is the long-take model. One generation runs to 30 seconds at any exact length, so a shot can match a narration line, and it has a 480p draft tier at 5 cents a second that Wan 2.6 does not expose. The Wan 3.0 model page covers it.
Wan 2.6 keeps two things over both: the lowest Wan rate at 720p with sound, and a match for any library already rendered on it, where the next clip has to look like the last one.
First frame: what it does, and what it does not
A first frame pins the opening image exactly, and the clip animates out of it. A product photo or a generated still keeps its composition, and the move you prompt plays out from there. The first frame post walks through the technique and the image to video generator covers it from the still's side.
There is no last frame, so a shot cannot be pinned at both ends; Wan 2.7 and Wan 3.0 both take both keyframes. And there are no reference images on this endpoint: a still wired into the node's reference input is used as the opening frame instead, rather than carrying the subject into a new scene. For that, use a model with reference input, such as Wan 2.7 or HeyGen Video.
About the audio
Wan 2.6 generates ambience and effects with every clip, and the Video Generation node shows a Generate audio switch for it. Leave it on for b-roll that has to feel like a place. Turn it off when a narrator or a music bed will sit on top.
Narration is added in the graph: a Text to Speech node handles the voice, an Audio Generation node handles music or effects, and the Sequence node mixes them over the picture and can duck the clip's own audio under a voice. The AI voiceover generator covers the narration side.
What you actually pay on Treza
Wan 2.6 through a Treza pipeline is metered from a prepaid credit balance. One credit equals $0.01, credit packs start at $5, credits never expire, and a failed generation charges nothing.
Before a run starts, the platform estimates the media cost of the graph and refuses to start one your balance cannot cover. For Wan 2.6 the estimator holds the first-frame rate for the resolution you pick: 10 cents of provider cost per second at 720p and 15 cents at 1080p, and with no resolution set it holds the 1080p figure. A clip made from a prompt alone settles at the plain rate, and the difference returns to the balance when the run finishes. Our test clip was held at 50 cents and settled at 40. Reserves are holds, not prices.
For a finished video costed end to end, narration and captions included, see the AI video cost breakdown.
Calling Wan 2.6 as an API without building the integration
Direct API access to a video model means credentials, a submit-and-poll job pattern, retries, moderation rejections, and somewhere to store the output.
On a Treza AI video pipeline, Wan 2.6 is a node on a canvas. Set the duration to 5 or 10, pick the aspect ratio and resolution, wire in a first frame if the shot starts on a still, place it next to a script, voiceover, or captioning step, and publish the whole chain as one versioned endpoint. Your application calls a typed /invoke route with JSON in and JSON out, or points an existing OpenAI SDK at the OpenAI-compatible route.
The swap is what pays off later. When the next model wins on price or quality, you change the dropdown and republish, and the endpoint your product calls does not change. Our hosted MCP server exposes the same pipelines as tools, so an agent in Claude or ChatGPT can render a brief and return the link.
Where Wan 2.6 earns its place in a graph
Three jobs where a short clip with sound at a low fixed price is enough. Vertical b-roll for a faceless channel, where a 10 second 9:16 clip at 80 cents opens a segment with its own ambience and the voice carries it. Establishing shots in an explainer, 5 seconds at 40 cents each, stitched by the Sequence node, which handles up to 12 shots. And short vertical cuts for Shorts and Reels, where 720p is watched on a phone and the sound is part of the hook.
The fastest way to learn what it costs for your own work is to run your real prompt at your real settings and read the charge. Open the AI video generator, pick Wan 2.6, and generate once.
Frequently Asked Questions
How much does Wan 2.6 cost per second?
$0.08 per second at 720p from a prompt and $0.10 from a first frame, $0.12 and $0.15 at 1080p. Those are the listed rates for the endpoint we run as of October 2026. A 5 second 720p clip from a prompt is 40 cents of provider cost, which is exactly what our test render billed. On Treza the charge comes from a prepaid credit balance, and a failed generation charges nothing.
Does a first frame cost more on Wan 2.6?
Yes. Starting from a still raises the rate by a quarter: 10 cents a second instead of 8 at 720p, and 15 instead of 12 at 1080p. The estimator holds the first-frame rate before a run and returns the difference when a prompt-only clip settles.
What resolution does Wan 2.6 render?
720p or 1080p. Our 5 second 9:16 test at 720p came back at 720 by 1280 at 30 frames per second, which matches the setting. We have not rendered 1080p yet.
How long can a Wan 2.6 clip be?
5 or 10 seconds, nothing in between and nothing longer. For exact lengths, Wan 2.7 renders any whole second from 2 to 10 and Wan 3.0 from 2 to 30. For longer pieces on Wan 2.6, stitch generations with the Sequence node, which handles up to 12 shots, or assemble them in the timeline editor.
Does Wan 2.6 generate audio?
Yes. Ambience and effects are generated with the clip, and our test file carried a quiet AAC track with the room's sound. The Video Generation node shows a Generate audio switch for it, so turn it off when a narrator or music bed will sit on top. The rate is the same either way.
Can I call Wan 2.6 through an API without an Alibaba account?
Yes. On a Treza pipeline, Wan 2.6 is a node you configure once and publish as a versioned HTTP endpoint, with a typed /invoke route and an OpenAI-compatible route. A hosted MCP server exposes the same pipelines as tools for agents. Generations are metered from prepaid credits that start at $5 and never expire.
