Mischief 1 Is a Wan 2.2 Adapter With a Neuron Story
A startup grew rat neurons on electrodes for four years and shipped a video model: three to five seconds, 16 fps, five cents for a 720p clip. The API schema says more about the speed claim than the announcement does.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

On October 30, The Biological Computing Co. put an image-to-video model on fal.ai. The announcement says the company has spent four years growing real brain cells on electrodes to understand how they compute, and that the findings now exist as a product. It is called Mischief 1, it costs from half a cent per second of video, and the tagline describes it as at least five times faster than the model it is built on.
I cannot check the biology. What I can check is the API: the schema, the billing unit, the defaults, and the two numbers the product page puts beside each other. Those tell you more about what you are buying than the word "neuron" does.
The product is ordinary, which is the interesting part
Mischief 1 takes a still image and a written action and returns a short clip. Three or five seconds, at 832×480 or 1280×720. It renders at 16 frames per second on a 4N+1 frame grid, so a five-second request comes back as 81 frames (5.0625 seconds of footage) and a three-second request as 49. You are billed for the seconds you asked for rather than the frames you received: $0.005 per second at 480p, $0.01 at 720p. Five seconds at 720p is five cents. A hundred drafts is five dollars.
The condition image can be up to 25 MiB and 100 megapixels. A JPEG or PNG under the default 2.5 MB byte budget is sent as it is, and anything larger or in another format is re-encoded, scaled down only as far as needed. Fitting is center_crop by default, which keeps your aspect ratio and cuts the edges, with resize there if you would rather stretch.
None of that is remarkable in late 2026, and that is the point. The pitch is not a new interface. It is the same image-to-video call with a different cost line and a story attached.
Two defaults that decide your results
The first is the frame rate. Wan 2.2's Ti2V-5B model card advertises 720p at 24 fps on a single 4090. Mischief renders 720p at 16. Five seconds at 24 fps is 121 frames; five seconds at 16 fps is 81. That is roughly a third less sequence to denoise before any architectural change is counted. A lower frame rate is a defensible choice for a cheap draft tool, and motion at 16 fps reads differently from motion at 24, which is a look you may or may not want. It also means "faster" needs a frame rate beside it before the number means anything.
The second is prompt_processing, which defaults to enhanced. With that on, your prompt is rewritten against the condition image and the clip shape before anything renders. The seed field fixes the render only, because the rewrite takes no seed. So if you A/B two prompts with the defaults untouched, you are not comparing two prompts, you are comparing two rewrites of them, and the second run will not reproduce the first. Set it to original when you are measuring, and leave it alone when you are just trying to get a shot.
The schema also says, in the prompt field description, that long and specific prompts steer Wan best. That is not a marketing line, it is someone documenting their own endpoint, and it tells you which habits carry over if you already work with Wan.
The speed claim and the two numbers
Numerama's write-up supplies the part that is not on the model card. Mischief is not trained from zero: it is Wan 2.2 with a small software layer, under 0.1% of the base model's weights, inserted between its compute blocks. The claimed outcome is equivalent quality, at least five times faster, with 80% lower compute. The product page shows 12 seconds of generation against 142 seconds for Wan 2.2 alone.
142 divided by 12 is 11.8, not 5. Both figures can be true if the headline was measured on a different prompt, resolution or GPU than the pair on the page, and there is no paper saying which. That is the whole problem in one line: the speed claim is the product, and the evidence for it is two integers with no stated basis for comparison.
This does not make the model bad. Distillation already makes Wan 2.2 several times faster by cutting the number of denoising steps, and nobody calls that a miracle, because it is the normal way video models get cheap. The open question is whether a neuron-derived layer does something distillation cannot, or the same thing with a better story. Numerama reaches the same conclusion, in less polite French.
What the neurons actually are
TBC was founded by two neurosurgeons, Alex Ksendzovsky and Jon Pomeraniec, and runs labs in San Francisco and Baltimore where rat neurons and human stem-cell-derived neurons sit on grids of electrodes. The described method: turn a computational problem, pixels for instance, into electrical stimulation, record what the cells do, and look for recurring patterns of computation in that activity. The patterns become equations, and the equations go into ordinary artificial networks.
So no rat neuron works when you call the endpoint. The cells are a research instrument, four years of measurement that produced priors someone then wrote into a layer of a diffusion model. That is a legitimate way to do research, and from the outside it is indistinguishable from a well-tuned adapter, which is exactly why the missing paper matters.
The company has raised $52M, says it has applied the method to Nvidia's Cosmos for robotics, and describes a longer-term goal of putting real neurons back inside the compute loop. It is the first to ship a consumer-facing model this way, though not alone in the field: Cortical Labs in Australia sells a biological computer built on human neurons, and FinalSpark in Switzerland rents time on brain organoids.
How I would test it
The number that matters to anyone paying per clip is not seconds per generation. It is cost per keeper.
Take five prompts from work you have already done, the ones where you know what a good result looks like. Set prompt_processing to original, keep resolution and duration identical, and run four clips per prompt per model on the same recorded sequence of seeds: one fixed seed reproduces the same render rather than giving you a second attempt. Then record three things per prompt: wall-clock time to the first clip you would actually use, how many seeds you burned to get one, and the price you paid per clip on that model, because two models at different prices cannot be compared on time alone. If none of the four is usable, record that as no keeper and move to the next prompt, because "this model produced nothing I would ship in four attempts" is an answer too. At five cents a draft, a model that renders in 12 seconds and needs four tries loses to one that takes 40 seconds and lands on the second. This is the same accounting I use for local shot times, where the useful figure is the cost of a shot rather than the speed of a model.
A cheaper version of the same test, if you do not want to pay for the comparison: run ten 480p clips of your own material at 2.5 cents each, and count how many survive at export size. Twenty-five cents buys a real answer about whether the output is usable, which no announcement can give you.
Where it lands
At 480p, five seconds costs two and a half cents, which puts this in draft-territory: cheap enough to try an action three ways before committing to a longer pipeline. The 16 fps cadence and three-to-five second ceiling also set the limits. This is a tool for shots that are one action long, where the first frame already carries the composition, and the same restraint applies as with any Wan derivative: describe the camera as a fact and give the clip one thing to do.
The honest reading is a Wan 2.2 variant with a smaller workload, a reproducibility trap in its defaults, a genuinely low price, and a biological claim that nothing published supports or refutes yet. That is enough to justify a twenty-five cent test. It is not enough to rebuild a pipeline around, and the number that would change my mind is not a speedup. It is cost per keeper, measured on prompts you care about, by someone with no press release to write. The same skepticism applies to any step-reduction claim that arrives with a multiplier attached: the multiplier is easy, the keeper count is the argument.