More takes.
12x the clips changes how a director works by staying in a flow state. You explore instead of rationing, and the best shot is usually the one you would not have paid for.
Testimonial / my tests and review of TBC mischief 1
What if better AI didn't require more compute?
A personal testimonial, with my own tests and review, by Justin Hackney, Co-founder and CCO, Wonder Studios
For the last decade, AI got better by getting bigger. More chips, more power, more data centers. It works. But nature found another way.
The human brain remains the most efficient computer we know.
TBC learns directly from the brain. They put living neurons on a chip, showed them the world, and listened. What brain cells taught them now makes a video model faster and cheaper to run.
I call that a green model: intelligence that comes from living systems. I put TBC mischief 1 to the test. Here's what it can do.
Real images and video are turned into patterns of electrical stimulation.
Living neurons on a multi-electrode chip respond. Every electrical spike is recorded.
TBC decodes that activity and creates algorithms to make AI models compute better.
Algorithms improve open-source video models, adding less than 0.1% to the model's size.
Faster video models on standard silicon chips. The neurons stay in the lab.
Process as described by TBC.
Each card is one 5 second 720p take. mischief 1 at its price on fal and OpenRouter, $0.01 a second at 720p; Wan 2.2 A14B at fal.ai's list price. Pictures are opening frames from my mischief 1 clips.
Source: TBC benchmark, 18 Sep 2026, on 8x NVIDIA H100 SXM 80 GB (Massed Compute). Image-to-video, 1280x720, 81 frames at 16 fps (5 s clip). Wan 2.2 I2V-A14B at vendor default settings vs TBC mischief 1. Median generation time per clip: 141.6 s vs 12.0 s on 8 GPUs; 1,002.3 s vs 74.3 s on 1 GPU. Excludes model load and video encoding.
How long a take takes
mischief 1 from my runs on the canvas. Wan 2.2 A14B at standard settings, about 3.2 minutes a clip. Played 24 times faster than real time.
Three weeks of my tests
13 Oct: 840 clips in one day for $42. On Wan 2.2 at list price, the same day costs about $340.
Clip counts from my Dream Canvas records. Cost at mischief 1's current price of $0.05 a clip.
Try something, see what works, make a change, and try again, without second-guessing whether an idea is worth the cost or the time.
173 clips created with TBC mischief 1, in two sets from my board: profile shots and top-down shots. Hover over any tile to watch. The opening frames were generated with an image model; all motion was generated with mischief 1.
To create our launch video, I used Dream Canvas, a custom-built app that combines AI agents and different types of inputs to help us experiment and iterate. With mischief 1, I could explore ideas almost in real time, without breaking my creative flow. Here's how it worked:
Rebuilt from my real Dream Canvas boards and real mischief 1 takes, not screen recordings. Finish times are compressed but in their real order; the cursor and picks are staged.
mischief 1 Wan 2.2 1 dot = 10 clips
12x the clips changes how a director works by staying in a flow state. You explore instead of rationing, and the best shot is usually the one you would not have paid for.
As TBC's approach reaches more advanced models, it could reshape how films and commercials are made, with faster, more efficient generation that finally keeps pace with imagination.
How we counted
Speed figures are typical times, not yet a matched test on the same inputs. TBC's own published figure is 5x, so we show both.