Why must an edited AI video keep the exact length of the source?

The upstream message is explicit: *Edited video duration (9.958s, rounded to 1st decimal: 10s) does not match config duration (7s)*. The validator compares the produced file against the duration in the request and rejects a mismatch — after the model has already done the work. Two to five minutes gone, and the only thing that changed was a number in the request.

The tolerance is a tenth of a second, and the comparison rounds. That is why a 3.167-second source declared as 3 seconds fails — 3.125 rounds to 3.1, not 3.0 — while 9.958 declared as 10 passes.

Two rules, applied everywhere a video can enter an edit. First, duration is never taken from the form when editing: a conversational edit inherits the parent generation's actual length, and a video-to-video edit is measured from the file itself.

Second, every source is trimmed to a whole number of seconds before it is sent — a hard cut at N seconds produces exactly N.000, which removes the rounding ambiguity entirely. That applies to uploads, to clips picked from your history, and to sources fetched by URL through the MCP tool alike. Sources shorter than about three seconds are rejected rather than padded, because below that the length no longer rounds to a legal value.

The price follows the same logic: you are billed for the effective duration of the result, not for whatever the form last showed. The final charge is always the amount recorded on the finished generation.

Editing is not the tool — extension is. Since September 2026 Omni Flash runs on the 1.1 generation, which has a separate *extend* task: pick a finished clip, choose Extend, and the model appends 3–10 seconds to its end, reading the whole source first. Chain it and the clip grows to 40 seconds. It is a different job from an edit, and it is the only way an Omni clip gets longer; the original 1.0 generation refuses the task outright with *This model does not support video extension*.

For a longer continuous take than that, use Veo 3.1 extensions: each adds 7 seconds to an existing clip and they can be chained up to 148 seconds. Or cut several clips together; conversational editing keeps the character and lighting stable enough across shots that the cuts hold up.