Generating a short AI video can mean waiting through repeated attempts before a usable clip emerges. The Biological Computing Co. (TBC) is trying to shorten that loop with an unusual source of inspiration: measurements of living neurons.
The company is launching TBC Mischief, an image-to-video model based on Wan 2.2, through fal.ai on October 30. TBC says its version generates video five times faster at 80% lower cost than the underlying model, while improving output quality. Those are company-reported comparisons; the announcement does not provide an independent benchmark or enough test details to establish how the gains would transfer across workloads.
The central distinction is how biology enters the system. Mischief runs as software on conventional AI hardware. Customers are not sending prompts to living cells or installing biological processors. TBC uses neuronal activity in its research process, then translates what it learns into software that modifies an existing model.
What TBC Mischief offers
According to the launch announcement, Mischief will be available through fal.ai’s image-to-video endpoint, with free access through November 2. The announced price is $0.02 per second of generated video. At that rate, 100 five-second clips would cost $10, before any other workflow expenses.
TBC also says Wonder Studios tested the model in a production workflow that generated 100 five-second clips in roughly six minutes. That is a reported batch result, rather than a measure of the latency every individual user should expect. Concurrency, resolution, sampling settings, and serving hardware can all change the relationship between batch throughput and the time needed for one request.
For creative teams, the practical question is whether a faster model produces more usable footage per dollar. A lower generation price matters less if a shot requires many retries, loses continuity, or fails to follow the input image. Evaluating Mischief therefore means comparing both speed and successful outputs under equivalent settings.
From neuronal recordings to a software layer
On TBC’s website, the company describes a sequence of encoding computational challenges into electrical signals, measuring how cultured neuronal networks respond, and extracting principles that can be implemented in software. Its commercial approach is a post-training layer around diffusion transformer blocks, which it says adds less than 0.1% to the underlying model’s parameter count.
That small addition is significant to the product’s premise: an existing model can receive a targeted modification without being replaced by an entirely new foundation model. The biological component belongs to discovery and development; the resulting inference system remains digital.
The company’s Neural Dynamics Adapter technical post gives a more concrete example of its earlier research. In that work, TBC recorded cultured neural activity using a high-density array with 4,096 electrodes, converted those recordings into numerical representations, and incorporated a lightweight adapter into a diffusion model. The study used Oasis, an interactive Minecraft video model, rather than the Wan 2.2 model underpinning Mischief.
That distinction matters. A research result in one interactive environment can help explain a method without proving the performance of a separate commercial product. The adapter post also describes limitations, including its narrow test environment and results that varied with denoising settings. It supports a plausible engineering direction, while leaving broader generalization to be demonstrated.
Biology and conventional optimization can work together
A separate TBC post, A Sense of Space and Time, explores another way to transfer observations from biology into video generation. Its research architecture combines local spatial processing with a small temporal memory, using neural measurements to guide how information is retained across frames.
The post is also explicit that its deployed speed improvements combined model changes with serving optimizations. These included caching previously processed context, reducing sampling work, and serving multiple sessions efficiently. Its findings concern an Oasis-based interactive video system, not a direct benchmark of Mischief.
This offers useful context for the launch: biological research can inform a model architecture while familiar software techniques still contribute to the final system’s efficiency. Separating those contributions would help developers understand which gains come from the model, which come from the inference stack, and which depend on a particular deployment.
A route into existing AI infrastructure
Mischief follows TBC’s September announcement of a collaboration with Amazon Web Services. That announcement outlined work involving AWS Trainium, Amazon SageMaker AI, and a path toward commercial distribution through AWS Marketplace for its neuron-derived software. The earlier collaboration concerned text-to-video work; Mischief’s launch focuses on image-to-video access through fal.ai.
The common strategy is to bring the research into infrastructure developers already use. An API-accessible model is easier to evaluate inside an existing creative pipeline than a system requiring a new class of physical hardware. It also gives users a practical opportunity to test the company’s claims against their own images, prompts, and acceptance criteria.
TBC, led by CEO and co-founder Alex Ksendzovsky and COO and co-founder Jon Pomeraniec, says it is also exploring newer video models, language models, and physical AI. Those are development directions, rather than capabilities established by this release.
What would make the launch meaningful
The most useful next evidence would be a reproducible comparison with Wan 2.2: identical hardware and resolution, disclosed sampling settings, latency and throughput measured separately, and quality assessed across a broad set of inputs. Energy measurements would also be needed to connect lower inference cost with claims about computational efficiency.
Mischief’s launch moves TBC’s approach from research descriptions toward a tool that developers can actually test. Its longer-term significance will depend on whether the advantages survive outside selected demonstrations and across different models. For now, the concrete proposition is a small software modification, informed by biological experiments, that aims to make an existing video model more economical to use.
