On October 30, 2026, The Biological Computing Co. (TBC) released mischief 1, a video generation AI model that applies findings from neuron research. It is built on the open-weight video model Wan 2.2 and is available through fal and OpenRouter. Generation costs $0.025 per 5-second clip at 480p and $0.05 at 720p. Following the plan to commercialize the technology with AWS that we reported in September, the base model and pricing are now clear, and a working generation API is available.
The distinguishing feature of mischief 1 is that it reflects neuron responses measured in the lab in its software. Living cells do not perform computation when a video is generated; the model runs on ordinary AI hardware. TBC claims it generates video five times faster than its base model while maintaining equivalent quality. However, the published pricing and specifications need to be evaluated separately from the speed and quality gains attributed to neuron research.
A 5-second video for 2.5 cents: pricing and generation specs
The public API specification for mischief 1 requires both a prompt describing the video's content and an image_url specifying the starting image. Rather than generating video from text alone, it adds motion to an input image, an approach known as image-to-video. The output is MP4 at 16 frames per second (16 fps). Resolution is 832×480 pixels for 480p and 1280×720 pixels for 720p, both in landscape orientation.
The pricing published by fal is based on the requested video duration, with a different per-second rate for each resolution. Calculating for up to 100 five-second videos gives the following.
| Resolution | Price per second | One 5-second video | 100 5-second videos |
|---|---|---|---|
| 480p (832×480) | $0.005 | $0.025 | $2.50 |
| 720p (1280×720) | $0.01 | $0.05 | $5.00 |
The 720p rate is double the 480p rate, and a 5-second 480p video costs 2.5 cents. Even 100 videos come to just $2.50. Note, however, that these figures are generation fees calculated by multiplying the published unit price by requested seconds and number of videos. They are not total production costs, which would include taxes, discounts, editing work, and so on.
The duration you specify in the API differs slightly from the length of the file actually produced. In mischief 1, the frame count follows a "4N+1" rule (where N is an integer), so requesting 3 seconds yields 49 frames and 5 seconds yields 81 frames. At 16 fps, the actual playback times are 3.0625 seconds and 5.0625 seconds, respectively. Billing, however, is based on the requested 3 or 5 seconds rather than the playback length of the output file.
For anyone producing large numbers of short test clips, this pricing makes it relatively easy to estimate the budget needed. On the other hand, when comparing performance with other video generation AIs, matching only the specified duration does not guarantee identical conditions. Matching the actual frame count and frame rate as well allows output volume to be compared on equal terms.
Neuron responses incorporated into software, running on standard AI hardware
According to TBC, the parameters added by mischief 1 amount to less than 0.1% of the whole base model. The company's proprietary software runs on ordinary AI hardware, and no living neurons are needed for generation. Neurons are used in lab experiments, and insights gained from analyzing their responses are built into the AI model. Because the approach applies biologically derived computational principles to software, users do not need to prepare special facilities such as cell culture equipment.
The background of this technology can be gleaned from TBC's earlier technical blog post on the "Neural Dynamics Adapter". In that research, the company used Oasis, a 600-million-parameter video generation model, to test a method of incorporating information from the responses of living neurons into an AI. Specifically, input data is converted into electrical stimulation and delivered to cultured neurons, and the responses measured with 4,096 electrodes are converted into numerical representations. Based on that information, a small additional module called an "adapter" was trained.
This adapter is inserted into the early layers of the Diffusion Transformer (DiT) used for video generation. The part that converts neuron responses into a form usable inside the model is also trained, and the base model also receives limited additional training. Information is obtained from living neurons during the experimental stage, but inference after training is carried out entirely on computers. In other words, rather than using biological neural circuits themselves as computing devices, the approach applies insights from their behavior to an existing AI model.
That said, this was earlier research using Oasis, and the internal structure of mischief 1 is not necessarily the same. TBC has not disclosed the detailed architecture of the commercial model, so it is unclear whether the adapter's connection points and parameter count match those in the Oasis research. The research-stage configuration cannot simply be treated as the specification of the commercial model.
Wan 2.2 itself also comes in several models. The official GitHub repository publishes models such as "I2V-A14B," which generates video from images at 480p and 720p, and "TI2V-5B," which accepts both text and image input. However, TBC has only disclosed that mischief 1 is built on Wan 2.2 and has not specified which model it adopted. It is therefore impossible to tell whether mischief 1 is based on the 14B or the 5B model.
In addition, a small number of added parameters does not necessarily mean reduced computation. Generation time and power consumption vary with factors such as the number of processing iterations needed at inference and the volume of data transferred between memories. The fact that neuron research findings were incorporated into the software does not by itself prove how much efficiency improved.
Verifying the "5x faster" claim requires unified comparison conditions
TBC says mischief 1 can generate video five times faster while maintaining quality equivalent to Wan 2.2. However, it has not disclosed which specific Wan 2.2 model was used for comparison, the hardware, or the numerical precision. The number of videos processed simultaneously, the number of denoising iterations, and quantitative quality evaluation results have also not been fully published. The 5x figure is therefore, for now, TBC's announced value, not a performance gap confirmed by independent verification.
The official Wan 2.2 repository introduces existing acceleration techniques, including step distillation and quantization, as well as sparse attention that narrows the computation targets and caching that reuses calculation results. To determine how much the design derived from neuron responses contributed to the speedup, it would need to be compared with applying these conventional acceleration methods to the same base model. Whether generation speed improved and whether that improvement arose from the biologically derived technology must be verified separately.
Users comparing performance themselves also need to pay attention to API settings. mischief 1 provides a seed to control the randomness of results, but under the default prompt_processing=enhanced setting, the entered description is rewritten before generation. The random seed is not applied to this rewriting, so even with the same image, description, and seed, the description passed to the video generation model may not be identical. Fixing the seed alone cannot fully align comparison conditions.
To unify input conditions, you need to specify prompt_processing=original, which does not rewrite the prompt, and also fix the random seed. It is also advisable to save the starting image and original description and to keep the resolution and requested duration the same. Image handling differs as well: center_crop crops while preserving the aspect ratio, whereas resize stretches the image to fit the specified resolution. Mixing the two changes the content and shape of the input image, so the same setting should be used when comparing.
Still, these are basic measures for aligning comparison conditions and do not guarantee that exactly the same video will be generated after a service update. To verify speed, you need to record the base model, hardware, and numerical precision used, and unify resolution, fps, frame count, number of simultaneous jobs, and number of denoising iterations. Image quality and video consistency for the same inputs must then be evaluated as well. Simply cutting generation time by reducing frame counts or iterations does not justify concluding that the neuron-derived technology contributed to the speedup.
The earlier research using Oasis also includes an example where evaluation results changed with generation conditions. In TBC's technical blog, 10 evaluation videos with different starting scenes and motions were compared using entropy as a proxy for information retention. While the research as a whole reports improvement from the adapter, under conditions using a generation method called DDIM with 40 steps, the base Oasis model slightly outperformed the model with the adapter added in both image quality and information retention. The company also acknowledges that the effect on energy efficiency requires quantitative verification going forward.
This is not a result that negates the performance of mischief 1 itself; it is merely one condition in earlier research using a different model. At the same time, improvements confirmed in earlier research do not necessarily carry over to the commercial model. Results from company-published technical blogs and product announcements should be evaluated separately from peer-reviewed research and independent third-party verification.
Cheap generation is not the same as the cost of videos you can actually use
OpenRouter's model page lists use cases for mischief 1 such as exploring ideas, previsualizing footage, creating mood boards to share atmosphere, and prototyping multiple variations. For such uses, generating many candidates and comparing and selecting among them matters more than producing one highly polished video from the start. A low generation price lets the same budget cover more compositions and motions, broadening the choices in the production process. With mischief 1's published pricing, the cost of this kind of iterative prototyping can now be calculated concretely.
However, even if generation fees are low, production cost varies greatly depending on how many usable videos you actually get. For example, if you generate n five-second videos at 480p, the generation fee is n × $0.025. If the proportion of videos suited to the purpose is the adoption rate p, the number of usable videos is n × p. Dividing the generation fee by the number of adopted videos gives the average generation fee per usable video.
For 5-second videos at 480p, the generation fee per usable video is $0.025 ÷ p. Likewise, at 720p it is $0.05 ÷ p. Here p is a ratio greater than 0 and at most 1, and it is not a measured adoption rate for mischief 1. If not a single video can be adopted, the per-adopted-video cost cannot be calculated, and the labor costs of selecting and correcting generated videos are not included in this calculation.
To evaluate actual production costs, you need to set adoption criteria for each use and measure the share of videos that meet the required motion and image quality. By saving generation conditions and input data and repeating trials, you can compare the cost per usable video. If TBC were to publish speed and quality comparisons under identical conditions, it would also be easier to judge how much efficiency gain the neuron-derived technology actually delivers. How much value mischief 1's low price holds in real production settings will depend not only on the generation unit price but on the adoption rate, quality, and verifiable performance differences.