Wan 2.7 puts the audio inside the render, and takes your editing seam with it

What a two-stage pipeline gave you, besides sound, was a seam you could edit. Score in one place, picture in another, and a client note about the narration costs one render of one asset. This model closes that seam: the guide describes audio synthesised inside the same generation, or synchronised against a file you supply ^[18]. The comparison is not one call against two. It is one call against two, multiplied by how often someone changes their mind about the audio.

Then the arithmetic against the ceiling. Fifteen seconds is the maximum, and the guide says plainly that this rules the model out for longer-form production ^[5][8]. A 60-second cut is four generations at minimum ^[14], each carrying its own generated track, so the joins in the picture are joins in the audio too. The supplied-audio path has a matching gap: sync is documented as working best on files of 3 to 30 seconds ^[7], and everything from 15 to 30 seconds in that band, half its width, has no single clip long enough to sit under it ^[15].

Cost per approved asset is where the pitch thins. The A/B workflow the guide recommends is ten variations of a product demo, with seed control for reproducibility ^[12], while the same document says a single five-second render at full quality needs significant compute and considerable time ^[9]. It quotes no price, no throughput and no benchmark ^[16], so any re-cost against an existing vendor stack has to be measured locally rather than read off the page.

"1080p with audio" is also a ceiling rather than a default. The guide flags quality variance on the Replicate-hosted 1080p path and notes the 1.3B variant is steadier at 480p ^[19]. On-screen text in Chinese or English comes out unpredictable, often blurry or distorted ^[1], and prompted camera moves can turn jittery ^[2]. Negative prompts reduce unwanted elements without guaranteeing their absence ^[10].

The upgrade advice inside the guide argues against the collapse story for anyone already running two stages. It says choose 2.5 if inference speed is critical, and 2.7 if you need reliable audio sync ^[3]. A team with a working text-to-speech or music step is being offered sync it may not be short of, and paying latency for it. The comparison against 2.6 is cut off in the material available ^[17], so treat the version ladder as unfinished rather than settled.

So the honest re-cost is narrow. Single-pass audio wins where the output is disposable and nobody downstream gets to request a different track: B-roll behind a stream, prompt variations nobody will revise, filler that gets replaced rather than edited ^[13]. Where the audio is an asset with its own review cycle, the second stage was buying control, and this removes it.