TBC Mischief reduces text-to-video costs with a biological layer
TBC launches Mischief, an image-to-video model available via fal.ai from October 30. The company claims to cut costs by five times and greatly speed up generation, but these gains remain to be confirmed outside its own tests.
Biological Computing Co. (TBC) is launching Mischief on October 30, an image-to-video model accessible via fal.ai and based on Wan 2.2. The company claims its system generates videos five times faster, at 80% lower cost, while improving quality. At this stage, these figures come from TBC and are not based on an independent benchmark. This topic matters to creatives as well as developers, as it affects usage cost, iteration speed, and how biological research can feed into conventional AI software.
Une couche inspirée de neurones, mais pas un ordinateur vivant
Mischief remains software running on conventional AI hardware. Clients do not send requests to living cells nor install biological processors in their studios. TBC explains that it uses neuronal activity as a research tool, then translates what it learns into a software layer applied to an existing model. The promise is simple to state but harder to validate: to do better with a slight modification rather than a new foundation model.
According to TBC's technical description, this approach takes the form of a post-training layer around diffusion transformer-type blocks. The company indicates the addition represents less than 0.1% of the total number of parameters of the underlying model. In other words, the graft is tiny on paper. This is often where the most interesting technical bets hide.
Ce que les utilisateurs verront dans fal.ai
According to the announcement, Mischief will be accessible in fal.ai's image-to-video endpoint, with free access until November 2. The announced price is $0.02 per second of generated video. Based on this, 100 clips of five seconds would cost $10, excluding additional workflow costs. For a creative team, the question is not only the sticker price but the number of retakes needed before obtaining a usable sequence.
TBC also cites a test conducted by Wonder Studios: 100 clips of five seconds were generated in about six minutes. Again, this is a batch result, not a guaranteed latency for each request. Competition, resolution, sampling parameters, and the server used can greatly change the outcome. A video fast on paper can become much slower when examined closely.
Des gains qui mêlent modèle et infrastructure
TBC's sources show that performance does not depend on a single lever. In another technical article, the company describes an architecture combining local spatial processing and reduced temporal memory, with service optimizations such as caching already processed context, reducing sampling work, and better management of multiple sessions. These choices concern an Oasis system, not Mischief, but they set the framework: gains often come from a combination of methods, not a single magic trick.
TBC also published an older example, Neural Dynamics Adapter, where the team recorded neuronal activity of networks cultured with a high-density matrix of 4,096 electrodes before integrating a lightweight adapter into a diffusion model. This study used Oasis, an interactive video model linked to Minecraft, not Wan 2.2. The difference is important: a lab result does not mechanically prove the performance of a distinct commercial product.
Pourquoi ce lancement regarde aussi du côté d’AWS
Mischief fits into a broader TBC strategy, which announced in September a collaboration with Amazon Web Services around Trainium, SageMaker AI, and potential distribution via AWS Marketplace for its software derived from neuronal research. The new launch goes through fal.ai, but the logic remains the same: to integrate into tools developers already use. This is more concrete than talk of the "next generation" and easier to test in a real production chain.
The company, led by CEO and co-founder Alex Ksendzovsky and COO and co-founder Jon Pomeraniec, also says it is exploring new video models, language models, and physical AI. These directions are still on the roadmap, not demonstrated capabilities from this release. For now, Mischief mainly serves as a demonstrator: a minimal software modification, powered by biological observations, meant to prove it really reduces usage cost without degrading results.
Ce qu’il faudra vérifier avant de parler de percée
The most useful elements are still missing to judge Mischief's real impact. A reproducible comparison with Wan 2.2 would be needed, on the same hardware, with the same resolutions, sampling parameters, and separate measurements of latency and throughput. Quality tests on a broad set of images would also be necessary, as well as energy measurements if one wants to link inference cost to measurable computing efficiency.
As it stands, TBC is bringing to market a tool developers can try immediately. This is already more solid than a lab promise. But between a nice demonstration and a lasting production advantage, the road is short on paper, never in practice.