TBC mischief 1: The AI Video Model That Learned From Living Neurons

TBC mischief 1 is an image-to-video AI model that generates short clips for half a cent per second, and its creator says the speed gains come from experiments with living neurons, not from the usual bag of optimization tricks. The Biological Computing Co. (TBC), a San Francisco startup, launched the model on October 30, 2026 through fal and OpenRouter.

The pitch is seductive: an AI model that runs five times faster than its base model while matching its quality, built on Wan 2.2, the open-weight video model from Alibaba. The price is strikingly low. At $0.005 per second, a five-second clip costs 2.5 cents, and one hundred such clips cost $2.50. That makes it, by TBC's accounting, the lowest-priced image-to-video model on both marketplaces.

But the story splits in two. There is the product, which is real and testable today. And there is the biological claim, that the efficiency was discovered in living neurons, which deserves a closer look before you repeat it at a dinner party. This article walks through both: what mischief 1 is, what it costs, what "neuron-derived" actually means, and which of the company's claims you should hold lightly.

Contents

• What mischief 1 is — and what it isn't
• The price math, worked out
• The neurons stay in the lab
• Three claims to hold lightly
• How it compares to other AI video options
• Who should try it (and who should wait)
• Pro tips and common mistakes
• FAQ

What mischief 1 is — and what it isn't

mischief 1 is an image-to-video model: you upload a starting image, add a text prompt, and it generates a short video clip. The listings support three- and five-second clips in 16:9 widescreen, guided by your input image. It runs on conventional AI hardware, your GPU, fal's servers, OpenRouter's inference stack, the same as any other model.

What it is not: a biological computer. It is also not a from-scratch foundation model. TBC took Wan 2.2, an existing open-weight model, and applied what it calls a post-training software layer around the model's diffusion transformer blocks. That layer, the company says, adds less than 0.1% to the base model's parameter count: a tiny add-on, not a new model trained from neurons up.

That distinction is the whole story. TBC's commercial approach is to keep the base model and add a small modification shaped by biological research. If the approach works, it is a new source of efficiency. Most AI speed gains today come from quantization, distillation, and caching, techniques the whole industry shares. TBC says its optimizations were derived from direct measurements of living neurons, not merely inspired by them, which would make them hard for competitors to replicate. That is a big claim, and it is currently backed by company descriptions rather than peer-reviewed publications.

The price math, worked out

Pricing is where mischief 1 is hardest to dismiss. On OpenRouter the listing shows $0.005 per second at 480p and $0.01 per second at 720p. The arithmetic:

• One 5-second clip at 480p: $0.025
• One hundred 5-second clips at 480p: $2.50
• One 5-second clip at 720p: $0.05
• Two hundred 5-second clips at 480p: $5.00

For UK and Canadian readers: both marketplaces bill in US dollars. At October 31, 2026 exchange rates (roughly £0.75 and C$1.40 to the dollar), the $0.005 rate works out to about £0.0037 or C$0.007 a second. A five-second 480p clip costs around £0.02 or C$0.035, and one hundred clips about £1.87 or C$3.49. Your card provider converts at its own rate, so treat these as close approximations, not quotes.

For context on how far video prices have fallen, Google's Veo 3.1 replacement is priced around $0.10 a second, roughly 20 times what mischief 1 charges at 480p. Budget creators have options: our roundup of the best Sora alternatives maps the affordable end of the market, and Kling 4.0 offers longer 30-second generations for different use cases.

The fine print: pricing on marketplaces shifts, and TBC's "lowest-priced image-to-video model" claim was made at launch. Both platforms list competing models with different settings, resolutions, and promos, so treat the ranking as a snapshot, not a permanent title. Also note that 480p is fine for storyboards, drafts, and social cuts; it is not what you'd ship in a finished commercial.

The neurons stay in the lab

Here is what TBC says happens in its research process. Scientists grow living neurons on electronic plates fitted with electrodes. They send electrical signals to the cells, then measure how the cells respond. The researchers hunt for patterns in those responses that could make AI models more efficient, translate the promising findings into algorithms, and test those algorithms in existing AI models.

The company's own summary is worth quoting directly: "The neurons stay in our lab. What leaves is cheaper, faster, and better AI." Customers receive software that runs on standard AI hardware. Nothing biological leaves the building.

This matters because the marketing frame, "AI powered by living brain cells", invites a different picture: cultured neurons churning through your video prompts. That is not what happens. TBC's approach differs sharply from, say, Australia's Cortical Labs, which builds systems where living neurons actually perform some of the computation. TBC uses neurons to develop ideas for digital algorithms. The research process involves biology; the product does not.

There is also a research lineage worth noting. TBC published a technical post about its Neural Dynamics Adapter, describing how it recorded cultured neural activity with a high-density array of 4,096 electrodes, converted the recordings into numerical representations, and added a lightweight adapter to a diffusion model. That work used Oasis, an interactive Minecraft-style video model, not Wan 2.2, and not mischief 1. It supports the plausibility of the method, but it does not independently prove the 5x claim for the shipped product.

Three claims to hold lightly

TBC is a startup with a story to sell. Three of its central claims need a second source before you quote them as fact:

1. "5x faster than its base model, matching quality." This is a company-reported comparison. The launch announcement provides no independent benchmark and not enough testing detail to show how the gain transfers across resolutions, prompts, and scenes. A reproducible test, same hardware, same resolution, disclosed sampling settings, does not exist yet. TBC's own research post is also explicit that its deployed speed improvements combined model changes with serving optimizations like caching and reduced sampling work. Until someone tests mischief 1 against Wan 2.2 under identical conditions, treat 5x as the company's number, not an established fact.

2. "Lowest-priced image-to-video model on both marketplaces." Again a company claim, and one that depends on the current listings. The $0.005/second price is real and verifiable on OpenRouter, two independent outlets confirmed it, but the ranking against every competing model is TBC's assertion, not an audited comparison.

3. "World's first publicly available neuron-derived AI video model." "First" claims are marketing until proven otherwise. More importantly, the word "derived" does a lot of work here: the biological research contributed to a small software layer, and the underlying model is a conventional open-weight one. The claim is defensible in a narrow sense, nobody else is selling a video model built this way, but it does not mean biology does the computing.

None of this makes mischief 1 a bad product. It makes it a launch-day product: interesting, affordable, and not yet independently validated.

How it compares to other AI video options

The table shows the real choice. If you need long generations, Kling 4.0's 30-second ceiling wins. If you need the highest quality per clip, premium models still lead. If you need to generate a lot of short clips cheaply, storyboarding, testing takes, rapid iteration, mischief 1's price is in a different league. And if you want zero marginal cost and own a GPU, Wan 2.2 itself is still there as open weights.

Who should try it (and who should wait)

mischief 1 fits creators who iterate in volume: storyboards, animatics, social-media drafts, ad variations. At 2.5 cents a clip, the cost of trying ten takes instead of two drops to pocket change, which is exactly the point Wonder Studios co-founder Justin Hackney made in TBC's announcement, describing how cheaper iteration let his team "push the work much further within the same time."

It also fits developers building video features into products, where per-second pricing compounds across thousands of users. A model at $0.005/second makes features affordable that would be hard to justify at $0.10/second.

Who should wait? Anyone whose workflow depends on the 5x speed claim or on output quality matching Wan 2.2 exactly. Those claims are unvalidated. If your pipeline has strict latency budgets or quality gates, run your own comparison first. The price makes the test cheap to do.

Pro tips and common mistakes

Test speed yourself, the cheap way. Because the price is negligible, run the same image and prompt through mischief 1 and Wan 2.2 (self-hosted or via another marketplace) and time both. That gives you a real number for your workload instead of the company's number for theirs.

Don't confuse 480p with finished work. The headline $0.005 rate is 480p. The 720p rate doubles it. Decide which tier your deliverable needs before budgeting. A storyboard month at 480p and a client month at 720p are different budgets.

Mistake: repeating the biology story uncritically. "Trained on living neurons" is not what happened. If you write about or demo this model, the accurate frame is "efficiencies derived from neuron research, shipped as a small software layer." Getting this wrong undermines your credibility with technical audiences.

Mistake: assuming the price ranking is permanent. Marketplace prices move. Check the fal and OpenRouter listings before quoting "cheapest" in anything you publish.

Watch the AWS track separately. On September 23, 2026, about five weeks before this launch, TBC and AWS announced a collaboration: the neuron-derived model would run on AWS Trainium chips, deploy through SageMaker AI, and reach enterprise customers through AWS Marketplace. That is the enterprise distribution side of the same platform. This week's fal and OpenRouter launch is the self-serve, creator-facing side. Don't mix the two. The AWS track has separate pricing and, as of now, no general-availability date.

FAQ

What is TBC mischief 1?

mischief 1 is an image-to-video AI model from The Biological Computing Co., launched October 30, 2026. It turns a starting image plus a text prompt into 3- or 5-second clips in 16:9. It is built on the open-weight Wan 2.2 model and is available through the fal and OpenRouter marketplaces.

How much does TBC mischief 1 cost?

On OpenRouter it costs $0.005 per second at 480p and $0.01 per second at 720p. A five-second clip costs 2.5 cents at 480p or 5 cents at 720p; one hundred five-second clips cost $2.50 at the lower resolution. UK readers can reckon roughly £0.02 per five-second clip and Canadian readers about C$0.035, at October 31, 2026 exchange rates. Pricing can change, so confirm on the live listing.

Does it use living brain cells to generate videos?

No. Living neurons were used in TBC's research process: grown in the lab, measured, and studied for computational patterns. The shipped product is conventional software that runs on standard AI hardware, with a small add-on layer TBC says adds less than 0.1% to the base model's parameters.

Is the 5x speed claim proven?

Not independently. TBC says mischief 1 generates video five times faster than its base model at matching quality, but no independent benchmark or testing methodology has been published. TBC's own research posts note that deployed speed gains combined model changes with serving optimizations, so the neuron-derived contribution is unproven in isolation.

How do I use mischief 1?

Through the fal listing or the OpenRouter listing, where you upload a starting image, add a prompt, and receive a short clip via the marketplace API. TBC also offers a creation app for non-developers.

Is mischief 1 open source?

No. The base model, Wan 2.2, is open-weight, but TBC's neuron-derived layer is proprietary software delivered through fal, OpenRouter, and TBC's own app.

On October 31, 2026, mischief 1 is a remarkably cheap way to make short AI clips with an unusual origin story. Whether the biology is a breakthrough or a marketing frame, the product itself is testable for 2.5 cents a try, and that is a low bar for curiosity. If you work with video, compare it against the current alternatives and run your own numbers.