Introduction

There's a new video model out, and its design started with real brain cells. The Biological Computing Co. (TBC), a San Francisco startup, launched mischief 1 on October 30. It's an image-to-video model that the company calls "the world's first neuron-derived AI video model." It's API-only, and you can run it right now on fal (tbc/mischief-1/image-to-video) and OpenRouter (tbc/mischief-1). It starts at $0.005 per second of video at 480p.

Despite the headline, no neurons run your clips. TBC grows living neurons on electrode arrays, stimulates them with signals derived from video, and records how they respond. It then turns what it learns into ordinary software. The neurons stay in the lab. The model runs on standard GPUs.

What shipped

  • Model: mischief 1, image-to-video (first frame plus a text prompt)
  • Vendor: The Biological Computing Co. (TBC). CEO Alex Ksendzovsky, COO Jon Pomeraniec
  • Modality: image-to-video, 16:9, 3- or 5-second clips at 16 fps, 480p (832×480) or 720p (1280×720)
  • Open vs API: closed, API only, with no weights released. It sits on top of the open-weight Wan 2.2
  • Where to run it today: fal and OpenRouter. TBC says AWS Marketplace is coming soon
  • Concrete fact: TBC's add-on adapter is under 0.1% of the parameter count. The company says it matches Wan 2.2 quality while running at least 5x faster and costing 80% less to run.

What "neuron-derived" actually means

According to TBC's launch release and technology page, the pipeline goes like this:

  • Living neurons are grown on electrodes and shown stimulation patterns derived from video.
  • Their responses show which visual features biological networks track efficiently.
  • TBC turns those findings into a small adapter that sits on an existing video model, here Wan 2.2.
  • Only the adapter ships. All inference runs on regular GPUs.

So think of it as biology-inspired model surgery, not wetware rendering. The speed and cost claims come from the company. No independent benchmark has been published yet. A Unite.AI write-up also gave a different per-second price, so check the live rate card on fal or OpenRouter before you budget.

Price math

  • 480p: $0.005 per second, so a 5-second clip costs $0.025 and 100 clips cost about $2.50
  • 720p: $0.01 per second, so a 5-second clip costs $0.05
  • Billing follows the seconds you request. A "5 s" clip actually renders 81 frames (about 5.06 s), and you aren't charged for the extra fraction.

TBC calls it the cheapest image-to-video model on fal and OpenRouter. At these prices it's well suited to the "try 40 variations before lunch" stage of a project.

Knobs on fal

  • prompt (required): fal's own docs say long, specific prompts steer WAN best
  • image_url (required): your first frame, up to 25 MiB
  • resolution: 480p or 720p (default 720p)
  • duration: 3 or 5 seconds (default 5)
  • image_transform: center_crop (default) or resize
  • prompt_processing: enhanced (default) rewrites your prompt against the image, and original uses it exactly as you typed it. Use original together with a fixed seed if you need reproducible results.
  • Every prompt and image goes through TBC's policy screen before rendering.

Quick start (fal, Python)

import fal_client  # pip install fal-client ; export FAL_KEY=...

result = fal_client.subscribe(
    "tbc/mischief-1/image-to-video",
    arguments={
        "prompt": "Slow push-in on a rain-soaked neon alley at night...",
        "image_url": "https://example.com/first-frame.png",
        "resolution": "480p",
        "duration": 5,
        "prompt_processing": "original",
        "seed": 7,
    },
)
print(result["video"]["url"], result["seed"])

On OpenRouter, the same model is tbc/mischief-1. Video jobs there are asynchronous: you POST to https://openrouter.ai/api/v1/videos and poll for the result. OpenRouter lists a median latency of about 15 seconds.

A prompt to try

Launch-day samples lean into Halloween, since it shipped for "Mischief Night." Here's a long, literal prompt in the style fal recommends. Pair it with a moody first frame of a pumpkin patch at dusk:

Photorealistic pumpkin patch at blue hour, low fixed camera at ground level between curling vines and drifting mist. The large ribbed pumpkin in the center foreground stays the same size and position. Slowly, a carved jack-o'-lantern face glows to life on its front: triangle eyes, a small nose and a wide jagged grin, lit warm orange from inside, flickering gently. Fog rolls left to right behind it, distant pumpkins stay dark, a few fireflies rise. End on the glowing jack-o'-lantern holding still in the quiet field.

Why it matters

  • Cost floor: at half a cent per second, image-to-video becomes cheap enough for rough storyboards, variation sweeps and pre-viz loops.
  • A new kind of efficiency claim: most speedups come from distillation or quantization. TBC is betting that biological experiments can show where to cut. The company says it's already applying the method to newer open video models, with language and physical AI next.
  • Partners: creative studio Wonder Studios worked with TBC on the launch. TBC also announced an AWS collaboration about five weeks ago.

Limits to know

  • The quality, speed and cost claims are self-reported, and there's no public leaderboard entry yet.
  • Clips are short (3 or 5 seconds), landscape only, and top out at 720p and 16 fps.
  • No weights or local install. Underneath it's Wan 2.2, so expect Wan-style strengths and quirks.

Original Source

Launch release: PR Newswire, Oct 30 · technology: tbc.co/technology · run it: fal · OpenRouter

Cheap clips, strange origins, and plenty of room to play. Go make some mischief. —Aurelia ♡