news

TBC's mischief 1 claims it matches Wan 2.2 quality at 5x the speed, with no published measurement

The Biological Computing Co. put a Wan 2.2 derivative on fal and OpenRouter on October 30, and named neither the base checkpoint nor the hardware behind its speed claim.

By Priya Shenoy · 2 November 2026

A San Francisco startup called The Biological Computing Co. put an image-to-video model on fal and OpenRouter on October 30 with two claims attached: it matches the quality of the open-weight model it is built on, and it generates video five times faster. Neither claim arrives with a measurement, a named checkpoint, or a description of the hardware the timing was taken on.

What shipped on October 30

TBC mischief 1 takes a still image and a prompt and returns a short clip. It is built on Wan 2.2, Alibaba's open-weight video model, and it runs on standard AI accelerators rather than on anything biological. The company's framing is that four years of laboratory work recording living neurons on electrodes produced efficiency ideas that were then written as software and applied on top of an existing model. The neurons stay in the lab. What customers buy is a hosted inference endpoint on two marketplaces.

The Next Web puts the rate at $0.005 per second of generated video, which works out to about 2.5 cents for a five-second clip, and states plainly that TBC "has not published independent benchmarks." TBC's own release states no per-second figure at all. It says only that mischief 1 is the lowest-priced image-to-video model on both marketplaces.

For price context on the family mischief 1 descends from, here is what Alibaba's current flagship, Wan 3.0, costs across the platforms in our radar.

"Matches its quality" is an evaluation claim with no evaluator

The speed number is the headline, but the load-bearing word in the announcement is quality. A derivative that runs five times faster than its base is only interesting if the output holds, and parity is the part that requires an evaluator: a prompt set, a scoring method, a sample count, and ideally someone other than the vendor. TBC names none of these. The release offers a customer testimonial from Wonder Studios about being able to push work further in the same time, which is a statement about workflow economics rather than a measurement of output fidelity.

The 5x figure has the same gap. Wan 2.2 was released in more than one parameter size, and TBC's materials identify the base only by family name, so a reader cannot tell which checkpoint the comparison runs against. Nor is there a GPU, a batch configuration, an output resolution, or a numerical precision attached to the comparison. Speedups in this category usually come from step reduction, caching, or quantization, and the company itself contrasts its approach with that standard toolkit, which makes the absence of a like-for-like setup description more consequential, not less.

The price-leader claim is a snapshot of two catalogs

Being the cheapest image-to-video option on fal and OpenRouter is a statement about two third-party listings on a given morning. Both catalogs add and reprice models continuously, and TBC's release neither dates the comparison nor lists the models it was measured against. It is the kind of claim that is probably true at launch and unverifiable a week later, which is why a dated screenshot or a named comparison set matters more than the superlative.

What would settle the interesting question is unglamorous: the same image and prompt set run through mischief 1 and through a self-hosted Wan 2.2 checkpoint at matched resolution and step count, timed on identical hardware, with the clips published. Until then the honest summary is that a cheap Wan 2.2 endpoint exists, and that the speed and parity numbers behind it are the vendor's alone.

Sources: The Biological Computing Co., The Next Web.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.