TBC releases AI video model shaped by measurements of living neurons

Creators and developers can now test an AI video model informed by measurements of living neurons, without using biological hardware. The Biological Computing Co. launched Mischief 1 on October 30, 2026, through fal and OpenRouter. Built on Wan 2.2, the image-to-video model turns TBC’s lab research into software that runs on standard AI hardware.

A low advertised price, with conflicting coverage

TBC’s launch release lists generation at half a cent per second, or $0.005. That puts a five-second clip at 2.5 cents and 100 such clips at $2.50. The company also claims Mischief 1 is the lowest-priced image-to-video model on both marketplaces; that ranking is TBC’s claim, not an independently verified comparison.

There is a material discrepancy in launch-day coverage. Unite.AI’s launch-day article gives a price of $0.02 per generated second, equivalent to $10 for 100 five-second clips. TBC’s own published rate is lower, but the two sources do not reconcile the difference. Neither figure includes other expenses in a creative workflow.

The neurons stay in the lab

The biological component is part of research, not the service handling customer requests. TBC says it grows living neurons, measures how they process information, and identifies computational principles it can translate into proprietary software. Customers are not sending prompts to living cells, and they do not need to install a biological processor to use the resulting model.

TBC describes its commercial approach as a software layer added after training, around the model’s video-generation blocks. The company says it adds less than 0.1% to the base model’s parameter count—the number of learned values in the model. The proposition is a small modification to an existing system, rather than a replacement foundation model.

Earlier research offers a more concrete example of the method. TBC recorded cultured neural activity with a 4,096-electrode array, converted it into numerical representations, and incorporated a lightweight adapter into a diffusion model. That work used Oasis, an interactive Minecraft video model, not Mischief’s Wan 2.2 base. Its narrow environment and results that varied with generation settings limit what it proves about this release.

Faster generation still needs a fair comparison

TBC says Mischief 1 matches Wan 2.2’s quality while generating video five times faster. Those are company-reported results. Unite.AI notes that the announcement lacks an independent benchmark and sufficient test details to establish how the gains would carry across different workloads.

There is a production-workflow example: TBC says Wonder Studios generated 100 five-second clips in roughly six minutes. That measures a batch of work, not the wait every user should expect for one clip. Simultaneous requests, output resolution, generation settings and serving hardware can all affect the relationship between total output and individual waiting time.

A useful comparison would therefore examine more than the advertised generation rate. Unite.AI identifies three practical checks for judging whether the efficiency claim translates into better creative work:

• Compare models on identical hardware and resolution, with disclosed generation settings rather than an unexplained headline speed.
• Measure single-request waiting time separately from the number of clips a system can produce in a batch.
• Count usable footage, including retries, visual continuity and adherence to the input image—not just seconds generated.