One Line In, Fifteen Models Working, You Nodding at Every Step

DomoAI announced InkoAI Agent on November 13. The pitch is one sentence: type an idea, get a short video.
The mechanism is the part worth reading. The agent runs a chain of tasks on its own, writes the prompts, sets the format and length for the platform you are aiming at, a vertical clip for Shorts or Reels rather than a 16:9 scene, and then, before generating anything, shows you what it plans to make and waits.
That pause is the product.
What the chain actually does
A creator starts with a line of text. The agent turns it into prompts, decides the container, and presents a plan. You can also bring your own material: scripts, storyboards, product shots. Earlier outputs stay in the thread, so a shot can be reworked or animated without re-uploading the file that produced it.
Behind the chat are fifteen image and video models rather than one. The published list includes GPT Image 2.5, Nano Banana Pro, Seedance 2.5, MiniMax H3 and DomoAI's own 2.5. Four different vendors, one interface, and an agent deciding which one gets which job.
There are also Skills, DomoAI's name for preset routines you launch with one click: short dramas, anime-style videos, cartoon clips, general stories with spoken lines or voiceover. In those flows you approve the script, then the characters, then the individual shots.
The company says more than five million creators use its platform, up from more than four million in May. There is no separate price for the agent.
Routing is the hidden job
Fifteen models in one chat is not a feature list, it is a scheduling problem that the user never sees.
A character sheet wants an image model that holds a face across variations. A three-second insert wants a video model that is cheap and fast. A talking shot wants a model with native audio. A style frame wants the model that renders that style best, which is a different model for anime than for photoreal product work. Somebody has to make those calls per step, and doing it by hand is exactly the tedious part of AI video production today.
The agent's answer is to make the calls and then show its work before spending compute. Which is a reasonable design, and also an admission: if the routing were obviously right, there would be no need to ask.
The quality ceiling is therefore the router, not any single model. A chain is as good as its weakest handoff, and a pipeline that switches models between shots has to solve character consistency across that switch, which is the same problem every multi-model workflow has and no one has fully solved.
Why microdramas are the beachhead
The Skills list is not random. Short drama, anime, cartoon, story with voiceover. Those are the formats where volume matters more than any individual frame, and where the audience is on a phone with the sound on.
The market numbers explain the targeting. Sensor Tower estimates short drama apps were downloaded more than 850 million times in the first quarter of 2026, up 140% year over year, with in-app spending around $750 million in the quarter, up 20%. DramaBox and ReelShort each took close to $140 million. Southeast Asia, Latin America and India accounted for more than three quarters of downloads.
That is a business with a brutal content treadmill and a fixed cost per minute. An agent that can turn a one-line premise into a vertical episode, with approval checkpoints where a human fixes the story, is aimed at exactly the producer who needs forty of those a month.
For marketers the pitch is different and narrower: one product photo becomes several ads styled like user-generated content, ready to test. That is a well-worn path, and the interesting part is again the checkpoint, since brand teams cannot ship an ad they have not seen.
The approval chain is a bet on distrust
Autonomy and approval are opposites, and consumer AI video has spent two years learning that full autonomy produces slop at scale.
InkoAI's answer is to be autonomous in the boring parts, prompt writing, aspect ratio, model choice, assembly, and to stop at every point where a human would later ask who approved this. It is a workflow claim dressed as an agent claim.
The cost of that design is approval fatigue. A ten-shot video with a checkpoint before each generation is ten interruptions, and users who click approve without reading have built a slot machine with extra steps. The preset Skills partly address this by batching approvals into script, characters, shots. Whether three checkpoints is the right number is an empirical question the product will answer with usage data, not with a blog post.
There is also the older caution in this category. OpenAI launched its Sora app in September 2025, passed a million downloads within weeks, and announced in March 2026 that it would shut the app down, after concerns about deepfakes. The company gave no explicit reason. Consumer video generation is a market where distribution can arrive fast and policy risk can arrive faster.
What it costs, and what it hides
DomoAI hasn't published a separate price for the agent, which is the least satisfying sentence in the announcement and the one a producer needs most, because the entire question with a routed pipeline is who pays for the wasted attempts.
A single-model tool has a simple bill: generations times price. A fifteen-model orchestrator has a bill that depends on its own decisions. If the router picks an expensive video model for a shot that a cheap one could have handled, the user pays for the routing error and never sees the line item. If it picks a cheap model and the shot comes out wrong, the retry costs more than the model ever would have.
The approval checkpoints are the only place a user can intervene in that economics, which is a quiet argument for them existing. A plan shown before generation is also an estimate shown before generation, if the product chooses to display one.
The other hidden cost is time. Fifteen models means fifteen loading paths, and a chat that pauses for approval five times across a ten-shot video is a session measured in hours rather than minutes. The companies winning this category right now are the ones that treat wall-clock time as a first-class feature, not a byproduct.
What to watch
Three things, in order of how much they would change the category.
Whether the router gets good enough to be invisible. If the agent reliably picks the right model per shot, the fifteen-model list becomes a genuine advantage rather than a support burden.
Whether consistency holds across model switches. The hardest technical problem in a multi-model pipeline is a character who survives a handoff, and no amount of chat UI fixes it.
Whether the checkpoint model survives contact with volume. Approval is cheap at three decisions and expensive at thirty. The product that finds the right granularity wins the microdrama floor, and the microdrama floor is where the money currently is.
For now, InkoAI is a clean statement of where consumer video tools are heading: not a model, not a studio, an orchestrator that keeps a human in the loop at the exact moments the human is still necessary.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.