A biology-inspired video model is now live

The Biological Computing Co. (TBC) made its Wan 2.2-based image-to-video model, Mischief, available on fal.ai on October 30. The company says its version runs five times faster than the base model, generates output at 80% lower cost, and also improves output quality. These comparisons are based on the company’s own statement; no independent measurement or details of the test conditions have been shared.

How neuron recordings become software

What sets Mischief apart is how biology enters the system. The model does not run on living cells; it runs as software on ordinary AI hardware. Users are not sending requests to live neurons or building a biological processor.

During its research process, TBC stimulates lab-grown neuron networks with electrical signals, measures their responses, and transfers the principles it derives from those measurements into software. The commercial approach is a post-training layer added around diffusion transformer blocks. According to the company, this layer adds less than 0.1% to the base model’s parameter count.

This small addition is important for the product’s core idea. An existing model can be modified in a targeted way without replacing it with a completely new base model. Biology stays in the discovery and development stage, while the inference system runs entirely digitally.

Pricing and access

Mischief will be available at fal.ai’s image-to-video endpoint, with free access until November 2. The announced price is $0.02 per second of generated video. At that rate, one hundred five-second clips cost ten dollars, excluding other workflow costs.

The company reports that Wonder Studios tested the model in a production pipeline and generated one hundred five-second clips in about six minutes. This is a batch result, not the latency of a single request. Concurrency, resolution, sampling settings, and server hardware can change the relationship between batch speed and single-request time.

Limits of the evidence

TBC’s previously published Neural Dynamics Adapter work records neuron activity with a high-density array of 4,096 electrodes, converts those recordings into numerical representations, and adds a lightweight adapter to a diffusion model. However, that work uses the interactive Minecraft video model called Oasis, not the Wan 2.2 model underneath Mischief. A research result obtained in one environment does not prove the performance of a separate commercial product.

The company’s post titled “A Sense of Space and Time” also explicitly states that some of the speed gains come from the model change and some from serving optimizations. These include caching previous context, reducing sampling work, and efficiently serving multiple sessions. So while biological research can inspire an architecture, familiar software techniques also contribute to the final efficiency.

What it means for developers

Mischief’s launch moves TBC’s approach from research notes to a tool developers can actually try. The AWS collaboration announced in September follows the same line: a distribution plan through Trainium, SageMaker AI, and AWS Marketplace. The goal is to bring the research into infrastructure developers already use.

For creative teams, the real question is whether the faster model produces more usable footage per dollar. Even if the generation price is low, savings evaporate if a shot requires many retries, continuity breaks down, or the model fails to follow the input image. Evaluation requires comparing both speed and successful output rate under the same settings.

The most solid way to verify the company’s claims is to run a reproducible comparison against Wan 2.2 on the same hardware and resolution, with disclosed sampling settings. Latency and batch speed should be measured separately, and quality should be evaluated across a broad input set. TBC says that, led by CEO Alex Ksendzovsky and COO Jon Pomeraniec, it is exploring newer video models, language models, and physical AI; these are development directions, not capabilities proven by this release.

Source: Unite.AI