MengTo quotes Mischief video generation at $0.005 per second
MengTo, who worked on Mischief 1’s landing page, quotes generation at $0.005 per second. Hasantoxr reports availability through fal and OpenRouter, with at least five times the speed of its base model.

TL;DR
- Five-second clips start at 2.5¢: MengTo, who worked on Mischief 1’s landing page, quoted generation at $0.005 per second.
- TBC claims at least 5x the speed of its base model, according to hasantoxr’s launch post.
- Mischief 1 is available through fal and OpenRouter, with AWS Marketplace listed as coming soon in the announcement quoted by hasantoxr.
The half-cent rate applies to 480p; 720p costs twice as much. A fixed seed alone does not reproduce a take, because the service rewrites prompts by default. The browser demo offers three free videos without sign-in.
Half-cent video at 480p
Mischief 1’s draw is cheap iteration: 100 five-second clips cost $2.50 at the entry resolution. The published fal rates are:
- 480p: $0.005 per generated second, or $0.025 for five seconds.
- 720p: $0.01 per generated second, or $0.05 for five seconds.
MengTo said he was used to paying 5¢ per second, ten times Mischief 1’s starting rate. Billing measures the duration of the generated video, not the time spent waiting for it.
The model is also listed on OpenRouter, which describes it as an image-to-video tool for ideation, previsualization, moodboarding and variant testing.
Starting images and 16 fps
The fal API schema specifies a short-clip workflow:
- Input: a starting image plus a text prompt, both required.
- Resolution: 832 × 480 or 1280 × 720, both landscape.
- Length: three or five seconds.
- Output: an MP4 at 16 frames per second.
- Defaults: 720p and five seconds, making the default clip 5¢ rather than 2.5¢.
The frame grid makes the rendered clips slightly longer than requested: 49 frames for three seconds, or 81 for five. Billing still follows the requested duration.
Prompt rewriting and repeatable takes
Prompt enhancement is enabled by default. Before rendering, the service rewrites the prompt using the starting image and clip settings; if rewriting fails or is rejected, it falls back to the original text.
A fixed seed controls the render but does not control that rewrite. According to the API documentation, reproducible results also require prompt_processing: "original". Both the image and prompt pass through TBC’s policy screening before generation.
Neuron-derived software on ordinary hardware
Mischief 1 builds on Alibaba’s open-weight Wan 2.2. In its launch announcement, TBC says it matches the base model’s quality while generating video five times faster.
The company grows living neurons, measures how they process information, and translates discoveries into proprietary software. Customers run that software on standard AI hardware; TBC says the addition amounts to less than 0.1% of the base model’s parameter count.
The outside efficiency test discussed by aakashgupta concerned a separate Minecraft video model. He reported roughly 4.4x lower inference costs in that test and noted that independent Mischief 1 benchmarks had not been published.
MengTo’s four-week landing page
MengTo’s contribution was the launch website. He described using Opus 5.5 to develop Three.js concepts, with a recorded workflow showing the interactive page and 3D neuron visuals.
The project took four weeks and seven versions, he said in a reply to someone who apparently thought it took an afternoon. “Feedback loop cannot be underestimated, even during the age of ai.”