What the tools actually put out

Twelve of the 15 tools publish a longest single generation, and they disagree by three orders of magnitude: six seconds fixed on CogVideoX, 30 on Wan 3.0, 30 minutes per scene on HeyGen. Nine publish their aspect ratios and 14 publish a resolution. As of 2026-09-22.

Inclusion rule. Tools whose own site publishes output limits in English on a page a buyer can open without an account: a priced page, an API reference, a usage-limits page or a model card. A limit the vendor does not state on that page is recorded as not stated rather than taken from a blog post, a demonstration or a reseller listing. Order. Alphabetical by tool name.

1Length is the spec that decides the format

Put this table next to the placement requirements and most of the matching is arithmetic. A 5 to 6 second bumper fits in one generation on every row that states a length at all. A 60 second non-skippable splits the register in half: two calls on Wan 3.0, four on Runway and Grok Imagine, eight on Google Veo, ten on CogVideoX, and one on the three entries whose ceilings are written in minutes rather than seconds.

That spread is not about quality. Each join is a place where lighting, framing and a face have to match footage that was generated separately, and the number of joins is set by this column rather than by the edit. Of the three rows that clear a full minute in one call, two do it because they generate a performance driven by an audio track rather than a scene driven by a prompt, which is a different product answering the same brief; the third cuts a two-minute request into six-second shots itself.

2Aspect ratio is published far less often than resolution

Almost every row here publishes resolutions, because resolution is what gets billed or what a model card was written to report. Six rows publish no shape at all, because shape usually is not priced. For advertising that is backwards: a placement specifies shape far more strictly than it specifies resolution, and a 16:9 file cannot be fixed into a vertical placement by re-encoding.

Two rows publish shape in a form the rest of the column cannot be compared against. Runway lists pixel pairs such as 1280:720, which answers shape and size in one token; HeyGen gives a pixel range from 128 to 4,096 on each side, which describes a space of legal outputs rather than a ladder. Where a vendor states nothing, the table says so rather than assuming the tool can produce whatever is asked for.

3Where the figure was read changes what it promises

Four rows here were read from priced pages and eleven from documentation: API references, usage-limit pages and model cards. The difference is not cosmetic. A priced page states what a buyer pays for and therefore names only the levers that carry a price, which is why duration and resolution are the best-covered columns on those four rows and shape is missing from three of them.

A documentation page states what a parameter will accept, so it names exclusions, defaults and retention periods that no pricing table would carry: four seconds is unsupported on one MiniMax model, a Grok Imagine request receives audio unless it asks not to, Google removes generated video from its servers after two days, Wan 3.0 returns a URL valid for 24 hours. None of those figures costs anything, and each of them can end a campaign.

• Veo 3.1 and Veo 3.1 Fast
  8 seconds, 6 seconds, 4 seconds, with 8 seconds only if 1080p or 4k or using reference imagesGoogle AI for Developers, Veo 3.1 / recorded 2026-09-22
• Longest output stated anywhere in this register
  30 minutes per scene, with 50 scenes per videoHeyGen documentation, usage limits / recorded 2026-09-22
• Video length
  6 seconds, at 49 framesfrom the model card tableCogVideoX-5B model card / recorded 2026-09-22
• Duration parameter
  An integer in the range [2, 30] seconds, and input plus output must not exceed 30 secondsAlibaba Cloud Model Studio, Wan3.0 API reference / recorded 2026-09-22
• Gen-4.5 image-to-video
  Landscape 1280:720, 1584:672 and 1104:832, portrait 720:1280, 832:1104 and 672:1584, square 960:960shape given as a pixel pair rather than a ratioRunway API, input parameters / recorded 2026-09-22
• Aspect ratio parameter
  1:1, 16:9, 9:16, 4:3, 3:4, 3:2 and 2:3, with 16:9 the defaultxAI docs, video generation / recorded 2026-09-22
• Aspect ratios offered
  1:1, 4:3, 3:4, 16:9, 9:16, 9:21, 21:9from 9:21 to 21:9Hedra, Character-3 model page / recorded 2026-09-22
• Resolutions
  720p, 1080p and 4Kpriced per stepKling AI, developer prices / recorded 2026-09-12
• Resolutions
  540p, 720p and 1080ppriced per stepVidu, API prices / recorded 2026-09-12
• Resolutions offered
  480p, 720p and 1080pPika, pricing / recorded 2026-09-12

4Sources

The first column of every row links to the page that row's figures were read from, and the last column points at this site's sheet for it where one exists. Requirements for the slots those outputs have to fit are on the YouTube placements page and the TikTok specifications; the rule that keeps the two halves from being averaged together is on method, and the columns themselves are set out one at a time under axes.