The two rows that charge by the second of output
Two entries in this register publish a per-second basis. One charges every generation that way, at a stated rate per second for two model names. The other charges the compositing step that way, at one credit per second of finished output. On both rows length is money. As of 2026-09-22.
| What the pages state | Runway | SceneMixer |
|---|---|---|
| What is billed per second | Every generation | The composite past the clip ceiling |
| Clip ceiling | 15 seconds on Gen-4.5 | 30 seconds on the Wan 3.0 and Seedance 2.5 tiers |
| Cheapest way to iterate | At the bottom of the range | Inside one clip, before compositing |
Inclusion rule. Billing statements from each vendor's own page, with the clip figure from the same page. No rates are quoted and no currencies are converted. Order. What is billed first, then the length it applies to.
1Per-second billing changes the order of work
Where every second is charged, framing and lighting should be settled at the shortest usable length and the length raised once. Discovering at fifteen seconds that the framing was wrong costs three times discovering it at five.
On a row with a flat rate per request the opposite is true: iterate at full length, because a short test costs the same as a long one. This pair is the reason a general rule about drafting cannot be written for the whole register.
2The two pages charge for different parts of the job
One charges the generation. The other charges the assembly that happens once the clip ceiling is passed, so requests inside thirty seconds are priced by the tier and only the longer pieces meet a per-second line.
For a sixty second non-skippable that matters. On the first row the whole minute is four per-second charges; on the second it is two tier charges plus a per-second composite, and the arithmetic has a different shape.
3What neither basis tells anyone
Neither publishes how many attempts a usable take costs, and that is the number that decides a budget. A rate per second multiplied by a length gives a floor, not a forecast.
These sheets record the basis because it changes the shape of the estimate, not because it produces one. The attempt count is the figure no page in this register publishes, and it is the reason a media plan carries a contingency.
4The rest of the column behind these two
| Generator | Duration |
|---|---|
| Amazon Nova Reel | 6 s increments, up to two minutes |
| CogVideoX | 6 s, at 49 frames |
| Google Veo | 4, 6 or 8 s |
| Grok Imagine | 1 to 15 s |
| Hedra | 10 minutes |
| HeyGen | 30 minutes per scene |
| Kling AI | Not stated |
| LTX-Video | Frames, never seconds |
| MiniMax | 4 to 15 s |
| Pika | 5 s and 10 s, 20-25 s on Pikaframes |
| PixVerse | 1 to 15 s on version 6 |
| Runway | 5 to 15 s on Gen-4.5 |
| SceneMixer | Thirty seconds on the Wan tiers, fifteen on H3 |
| Vidu | Not stated |
| Wan 3.0 | 2 to 30 s |
Inclusion rule. Every generator this register carries, on the one column this pair disagrees about. Rows that state nothing keep their place and are marked hollow. Order. Alphabetical by generator name.
Why these two. Everywhere else in this register length is a parameter and cost is a separate question. On these two rows the duration field is the invoice, which changes the order the work should be done in.
- How generation is billedCharged per second of output, with gen4_turbo at 5 credits per second and gen4.5 at 12
- Beyond the clip limitEpisodes are composited from segments, billed at 1 credit per second of outputhow length is reached
5Sources
Read from the Runway pricing guide and input reference, and the SceneMixer price list, checked 2026-09-22 and 2026-09-12 respectively. The column is described on duration, and each entry has its own cell page at Runway and SceneMixer, and the rule that keeps surfaces and generators apart is on method. Nearby pairs: Veo and Wan 3.0 on retention, Nova Reel and CogVideoX.