The 11x Model Swap That Cut an AI Anime Episode to ¥4,300

CreativeEdge, a small Japanese studio that runs a monthly production meetup, published the receipt for its own AI anime. The numbers are unusually specific for this field: two models, one scene length, one success rate, and the resulting bill for a three-minute episode.

The headline is a tenfold cost cut. The more interesting finding is that the cut only works because a large share of the episode stops being generated video at all.

The bill, line by line

The studio prices generations through Magnific's Premium+ plan: ¥6,600 a month for 45,000 credits, with each generation fixed at eight seconds and 768p or 720p depending on the model.

• MiniMax H3 Max Turbo, 320 credits: 140 generations a month, roughly ¥47 each
• MiniMax H3 Max, 640 credits: 70 generations, roughly ¥94 each
• MiniMax H3, 1,051 credits: 42 generations
• Seedance 2.5, 3,520 credits: 12 generations, roughly ¥516 each

That last line is the whole argument. Seedance 2.5 costs about eleven times what the cheapest MiniMax tier costs per clip, inside the same subscription.

What three minutes actually costs

A three-minute episode is 180 seconds. If all of it were generated video, that is roughly 23 clips of eight seconds. Assume a 25% success rate, which is what the studio uses: four generations per usable shot, about 92 renders in total. The studio's own episode does not generate all 180 seconds, as the later section explains, so these are the all-generated baseline numbers, not the final bill.

• MiniMax H3 Max Turbo: about ¥4,318
• Seedance 2.5: about ¥47,497

Per usable second that is roughly ¥24 against ¥264. The same 180 seconds of finished animation, same subscription, same operator. One number is a hobby budget and the other is a small production budget.

The 25% figure deserves attention, because it is not pessimism about the models. It comes from a deliberate method. For a 45-second scene in their series, the studio used 12 reference images and kept the prompt short. Seedance 2.5 accepts up to 30 references, and the studio leans on that capacity instead of describing the shot in words. Their stated reason is blunt: the more a production depends on the prompt, the more it becomes a gacha, and every failed pull is money gone. References narrow the output space before you spend anything.

The catch: you buy a second pass

MiniMax H3 Max Turbo is not Seedance. The studio says the visual quality is rough enough that the raw output cannot be used, so every shot goes through a creative upscale, in their case Topaz Astra 2. That pass costs extra money and, more importantly, hands-on time.

It is still cheaper than Seedance, by a wide margin, because Seedance is priced so far above the alternatives that a second processing stage fits inside the gap. The trade is real though. Upscaling does not restore information, it invents plausible detail. In the studio's own comparison, the upscaler added teeth and nose structure that were not in the source frame. For a flat cel look that reads fine. For a face held in close-up across four seconds, check it frame by frame.

Where the expensive model still earns its price

The studio did not conclude that Seedance is overpriced garbage. It concluded that Seedance is for a different job.

Their live-action series needs heavy reference support and photoreal continuity, so Seedance stays. Their AI idol project moved entirely to Wan 3.0 and cut generation cost by more than half with no large visible difference in a 60-second clip. But the studio adds the limit immediately: Wan 3.0 is not enough for action scenes or highly dynamic expression. A conversation in a room, yes. A fight, no.

That is a routing decision, not a model verdict. Price the shot, not the project. The same logic shows up in our Wan 3.0 cost breakdown: the value is in assigning shots to lanes, not in picking one model for everything.

The third bucket: 30% of the episode is never generated

Here is the part that changes the shape of the budget. The studio estimates that around 30% of the episode is not video generation at all. It is vector animation built with Claude Code: pans and zooms across still frames, incident light, simple effects.

The origin is an experiment in having Claude build something close to Adobe Flash, or more precisely its ancestor FutureSplash Animator. Vector graphics carry their motion as data, so those shots need no image generation at all, and a flat cel-anime look keeps render load low. The tool exports MP4, so nothing has to render in real time.

Once a shot is a still image plus a camera move, the video model was never the right tool. Paying ¥516 for a slow push across a painted background is the kind of expense that looks normal until someone itemises it.

Limited animation is a budgeting skill, not a style filter

The studio has been trying to redefine limited animation for AI production, and the cost pressure forced the prototype. Episode 1 came out as full animation, which was not the plan. Their first 45-second test failed for the opposite reason: too much movement, no limited-animation feel at all.

The distinction matters more than it sounds. Limited animation is not "fewer frames" as a filter. It is timing with weight: holds, pauses, fast transitions out of stillness, movement that lands on a beat instead of flowing continuously. You cannot get that by sampling every third frame. The studio's automated limited-anime pass on one eight-second clip took about 30 minutes, which is fine when it runs in the background while you do something else, and useless if the timing decisions are wrong.

The part worth stealing, and the part to avoid

Steal the accounting. Cost per usable second, at a stated success rate, is the only video price that compares across models. Steal the reference-first method, the upscale pass as a fixed line item, and the vector bucket for shots that do not need motion synthesis.

Do not steal the fully delegated iteration. The studio built an automated pipeline that converted generated clips into limited animation with calculated easing, and it worked. Then they stopped it. The reasoning is worth quoting in substance: the trial and error belonged to the agent, not to them, so the process did not become an asset. Claude left behind around 600MB of working data from repeated failures and fixes, and none of it was the studio's own knowledge. For client work there is a second problem: when a client's legal team asks how a shot was produced, "the agent figured it out" is a weak answer.

What they do instead is feed their own archive to the model. More than 60 hours of 2024 livestream recordings, plus slides, test clips, prompts, failures and fixes, are now context for the AI. They call it the most successful move of the year for their flagship project. The archive is the asset. The model is a reader.

A budget you can copy

1. Count clips, not minutes. Divide runtime by your clip length and round up.
2. Assume four attempts per usable shot unless you have evidence for better.
3. Price the two cheapest credible models at that volume, not at a single generation.
4. Mark every shot that is a still with a camera move, and take it out of the generation budget.
5. Add an upscale pass as a real cost, in money and in hours.
6. Keep your own log of what failed and why. That log is the thing you can reuse next episode, and the thing no subscription includes.

The elevenfold gap between two models is not really a review of those models. It is a statement about what your shots are. If half of them are paintings with a push-in, the expensive model was solving a problem you did not have.