Video models

Mischief 1: a video model shaped by living neurons, at a cent a second

Mischief 1 is built on Wan 2.2, an open video model, and costs an eighth as much as Wan 2.2 on the same service.

For four years, The Biological Computing Co. has grown living neurons on electrode arrays to study how they compute. On 30 October, the day before Halloween, it released the first product of that work: Mischief 1, a model that turns a still image into a short video clip.

It is cheap. On fal it costs 0.5¢ a second at 480p and 1¢ at 720p, so a five-second clip at 720p costs 5¢. That is the lowest price of the 26 video models we track. The next cheapest, HeyGen Video, charges 1.8¢ a second; Veo 3.1 charges 40¢.

The catch is sound. Mischief 1 makes silent clips, while 19 of the 26 models make a soundtrack with the video, HeyGen Video among them, and their prices include it.

Try Mischief 1 for free

Price per second of video, every model we track, on 30 OctoberAt 720p, or the closest size a model offers. Dashed: Wan 2.2, the model Mischief 1 is built on. makes sound, silent.
Mischief 11¢
HeyGen Video1.8¢ (768p)
Seedance v1.0 Pro Fast2.06¢
Grok Imagine Video 1.5 Lite3¢
Kling v2.5 Turbo4.2¢
Kling v2.64.2¢
Veo 3.1 Lite5¢
Seedance v1.0 Pro5.15¢
Seedance v1.5 Pro5.18¢
Grok Imagine7¢
Seedance 2.0 Mini≈7.56¢
MiniMax H38¢ (768p)
MiniMax H3 Max8¢ (768p)
Wan 2.2 (on fal)8¢
Wan v2.610¢
Wan v2.710¢
Wan v3.0 Video10¢
Seedance 2.0 Fast≈12.1¢
Grok Imagine Video 1.514¢
Wan v3.0 Video Prime14¢
Veo 3.0 Fast15¢
Veo 3.1 Fast15¢
Seedance 2.0≈15.1¢
Seedance 2.5≈23.1¢
Kling v3.025.2¢
Veo 3.040¢
Veo 3.140¢

Source: fal list prices for Mischief 1 and Wan 2.2; ModelCompare prices (the providers’ list prices) for the rest, with sound where the model makes it. ≈ marks token-priced models, estimated per second.

Where the neurons come in

TBC, a San Francisco start-up run by two neurosurgeons, feeds inputs to neurons from rats and from human stem cells as electrical pulses and records how they respond. It turns the patterns it finds into math, and the math into a small software layer wrapped around each block of an existing model. The neurons stay in the lab: Mischief 1 runs on ordinary chips, and the layer adds less than 0.1% to the model’s size, the company says.

The model under that layer is Wan 2.2, Alibaba’s open video model. That makes for a clean comparison. fal sells Wan 2.2’s image-to-video at 8¢ a second at 720p and 4¢ at 480p. Mischief 1 costs an eighth of that at both sizes. TBC says it is also at least five times faster than Wan 2.2, with the same quality.

Two Halloween stills

We made two stills with an image model and asked Mischief 1 to bring each to life. Each five-second clip at 720p cost 5¢ and came back in about 12 seconds. Neither has sound.

“The candle inside the pumpkin flickers, the black cat turns its head toward the camera and blinks, mist drifts across the lawn” · 5¢ · 12 s
“The children walk toward the camera, the witch swings the pumpkin bucket, the ghost sheet sways, lanterns sway in the breeze” · 5¢ · 12 s

New accounts start with 10¢ of free credit, enough for two Mischief 1 clips. Upload a still, say what should move and press Generate.

Try Mischief 1 for free

How we did it

Prices are list prices per second of video on 30 October: Mischief 1 and Wan 2.2 on fal, the rest as priced on ModelCompare, at 720p or the closest resolution a model offers, with sound when the model makes it. Seedance 2.x is priced in tokens; we estimate a second from its token price. Mischief 1 is also on OpenRouter.

The stills were made with FLUX.2 [klein] 9B on fal; each clip is Mischief 1 at 720p and five seconds, from the prompt shown under it, on 30 October. Time is from sending the request to having the file.

TBC’s claims about speed, size and how it uses neurons are from its site and an interview with The AI Innovator. We have not checked them.