Mischief 1 AI Video: the neuron-derived model that animates a still for half a cent a second

Mischief 1 — officially “TBC mischief 1” — is an image-to-video model from The Biological Computing Co. (TBC), a San Francisco company that uses living neurons to improve AI models [official]. It launched on October 30, 2026, deliberately timed to mark Mischief Night, the day before Halloween [official], and it is live on the developer marketplaces fal and OpenRouter [official].

The model is built on Wan 2.2, a widely used open-weight video model [official]. You supply a starting image and a text prompt — both are required — and it returns a short MP4: 3 or 5 seconds, 16:9 landscape, at 480p (832×480) or 720p (1280×720), at 16 frames per second, with no audio [official]. The image sets the composition; the prompt says what moves. A prompt on its own, with no image, is refused [reported]. There is no text-to-video mode, no vertical 9:16 output, and no 1080p [reported].

The “neuron-derived” part needs one careful sentence, because the launch headline invites the wrong picture. TBC’s discovery platform grows living neurons in the lab, measures how they process information, and identifies computational principles that can improve AI; those discoveries become proprietary software that makes existing models more efficient [official]. The neurons stay in the lab. Customers use software that runs on standard AI hardware and adds less than 0.1% to the base model’s parameter count [official]. Nobody is sending prompts to living cells, and no biological hardware serves your request.

TBC describes mischief 1 as the world’s first publicly available neuron-derived AI video model [official]. That is a first-claim about the research lineage, not a benchmark result — keep it separate from the speed and price numbers below, which are checkable.

The price is the story. At half a cent per second, a five-second clip costs 2.5 cents and 100 clips cost $2.50 [official]. TBC calls it the lowest-priced image-to-video model on fal and OpenRouter [official] — a company ranking, not an independent one, but one independent tracker, ModelCompare, found it the cheapest of the 26 video models it tracks as of October 30, at 1¢ a second at 720p against 8¢ for Wan 2.2 itself on fal [reported]. The honest comparison is against its own base model: mischief 1 costs one-eighth of Wan 2.2 on the same marketplace [reported].

The intended job is drafting, not finishing. OpenRouter’s own listing frames it as a fast, affordable way to bring ideas to life, ideal for ideation, previsualization, moodboarding and variant testing [official]. Three to five silent seconds at 720p will not carry a finished video — but at 2.5 cents a take, the cost of being wrong about a shot drops to almost nothing, and that is the whole pitch: test a dozen motions, keep one, render the keeper elsewhere.

The speed claim is the part to watch. TBC says mischief 1 matches Wan 2.2’s quality while generating video 5x faster and at a fraction of the cost [official]. No independent benchmark has been published, and the announcement does not disclose which Wan 2.2 variant was compared, on what hardware, or at what precision [reported]. One independent tester, Treza, ran a 3-second clip from a product photo in 19 seconds and was billed exactly the listed 1.5 cents [reported] — a price check, not a speed benchmark. Until someone runs the same prompt through both models side by side, treat “5x faster” as a company claim, not a measured fact.

It already has a production tester and an enterprise tail. Wonder Studios, an AI-native creative studio, tested mischief 1 in its own production workflow, and its co-founder is quoted in the launch release on the value of cheaper iteration [official]. The launch also came five weeks after TBC announced a collaboration with Amazon Web Services to bring the model to enterprise customers [official], with distribution through AWS Marketplace listed as coming soon [reported]. The consumer price is the headline today; the enterprise route is where TBC says the platform goes next.

There are two routes, and they serve different people. The official and API route is for developers and anyone who wants the raw per-second price; the third-party tool route is for anyone who wants a clip in the browser without an API key. The model is the same either way — the price you pay is not.

Route 1 — the official and API route (fal, OpenRouter, TBC’s app). On fal, the model page is tbc/mischief-1/image-to-video, with a playground and an API section [official]. On OpenRouter it is listed as tbc/mischief-1, served by a single provider, with a measured P50 end-to-end latency of 12.00 seconds and 98.44% availability over the past three days as read on November 8, 2026 [official]. TBC also points users to its own creation app at app.tbc.co [reported], but that app’s pricing and free tier could not be verified [unknown]. Upload a starting image, write a prompt that says what moves, what stays still, and how the camera behaves, then pick 480p for quick idea tests or 720p for better quality, and 3 or 5 seconds [reported]. Two details from the fal schema are worth knowing before you pay: prompt processing defaults to “enhanced”, which rewrites your prompt, so switch it to “original” and fix a seed when you want repeatable results [reported]; and the downloaded clip runs slightly longer than the requested duration — 3 seconds yields 49 frames and 5 seconds yields 81 frames at 16 fps, so about 3.06 and 5.06 seconds of playback [reported].

Route 2 — the third-party tool route (no API key). PixGPT runs mischief 1 in the browser: upload a reference image, describe the motion, preview the result next to the input, and download the finished clip as an MP4 in 1280×720 [reported]. A 5-second 720p generation costs 5 credits, and you get free credits on sign-up, from 2 credits per clip [reported]. PixGPT states plainly that it is an independent third-party interface, not TBC, and claims no official partnership [reported]. ImageToVideoAIFree also runs it: 480p costs 5 credits for 3 seconds or 10 for 5, and 720p costs 10 for 3 seconds or 15 for 5, with the cost shown before generation [reported] — but it requires sign-in and account credits, and its daily free generator is a separate model [reported]. Rewind.ai offers a free tool page for the model with 2,500 free tokens a day and no sign-up, rising to 5,000 with a free account, and token packs from $5 [reported]; what one video run costs in tokens is not stated on the page [unknown]. Treza, a pipeline tool, charges the model’s own rate from prepaid credits (packs from $5, never expiring, failed generations free) and its own test billed exactly the listed rate [reported].

Whichever route you take, start at 480p and 3 seconds: at 1.5 cents a take it is the cheapest way to find out whether the motion reads, and the silent, landscape-only output means you should plan the music, voice-over, and any vertical crop in your editor afterwards [reported].

The per-second rates below are confirmed twice: TBC’s own launch release states the half-cent rate and the 2.5-cent five-second clip [official], and OpenRouter’s model page lists the same two tiers in its FAQ [official]. The per-clip figures are arithmetic from those rates.

Resolution	Per second	3-second clip	5-second clip
480p (832×480)	$0.005	$0.015	$0.025
720p (1280×720)	$0.01	$0.03	$0.05

One warning about the headline number: fal’s defaults are 720p and five seconds, so the default clip costs 5 cents, not 2.5 [reported]. Pick 480p explicitly if you want the half-cent rate. And remember these are generation fees only — they do not include editing, sound, or the cost of the takes you throw away [reported].

One conflict to note: a launch-day article by Unite.AI listed the price as $0.02 per generated second — four times the rate TBC published [reported]. TBC’s own release and the live fal and OpenRouter listings both show $0.005 at 480p [official], so the $0.02 figure looks like an error in that one article, but it is still circulating in reprints — check the live listing before you budget, not a news summary.

This is an early bet, not a proven search term. On November 8, 2026, Google autocomplete returned no suggestions for “mischief 1 ai video”, “mischief 1 ai”, “mischief 1 video”, or “mischief 1 ai video generator”, and the suggestions for bare “mischief 1” were all the 1985 movie Mischief [reported]. The model is nine days old and the tool pages are up, but nobody is searching for it by name yet.

The head term already has several dedicated tools. PixGPT, ImageToVideoAIFree, and Rewind.ai each hold a dedicated mischief 1 page, and Treza lists it as a model [reported] — so this page is a guide, not a tool pitch, and it does not expect to outrank the tools on their own name. The room is in the long tail: the price, the neuron story, the drafting workflow, the comparison against Wan 2.2.

The speed claim has no independent test yet. “5x faster at the same quality” is TBC’s own comparison, with no published benchmark, no test conditions, and no disclosure of which Wan 2.2 variant, hardware, or precision was used [reported]. The only hands-on numbers so far are price checks, not speed checks: Treza’s 3-second test took 19 seconds end to end and billed exactly the listed 1.5 cents [reported]. OpenRouter’s own live stats show a P50 end-to-end latency of 12.00 seconds and 98.44% availability over three days [official] — useful, but latency is not a speed comparison against Wan 2.2.

The format ceiling is low and fixed. Three or five seconds, 16:9 landscape, 720p maximum, 16 frames per second, silent [official]. No vertical video, no 1080p, no audio or dialogue, no text-to-video [reported]. If your deliverable is a 9:16 social clip with sound, mischief 1 is at most the first draft — you will crop it, extend it, or re-render the keeper in a longer model.

The neuron story is a research story, not a product feature. Nothing about the living-neuron research changes what you can do with the model today: it is Wan 2.2 plus a small software layer, running on ordinary hardware [official]. Judge it on the price and the clips, not on the biology — and do not repeat the headline as “AI powered by brain cells”, because that is not what ships.

Three prices could not be verified. TBC’s own app at app.tbc.co exists and is linked from TBC’s posts, but its pricing and free tier could not be verified [unknown]. Rewind.ai’s per-run token cost for a video is not stated on its tool page [unknown]. And the $0.02-per-second figure in one launch-day article conflicts with TBC’s published rate and is most likely an error, but it has not been corrected on the record [reported].