Kimi K3.1 Pro for Video Scripts: API Pricing and a Real Pipeline Test
Kimi K3.1 Pro is Moonshot AI's hosted flagship, priced level with Claude Sonnet 5 and carrying a million-token context window. On our script probe it wrote a 30-second Short's script for about a sixth of a cent in 5.1 seconds. Here is the rate card, the same prompt on two Claude models for comparison, the fact every model got wrong, and how it fits a pipeline that writes, narrates, renders, and publishes.
Kimi K3.1 Pro is the hosted flagship of Moonshot AI's Kimi line. It reached OpenRouter on 23 October 2026 and joined the model picker on Treza's Text Generation node on 27 October, the first Kimi model in the node. There are no public weights for the Pro id, so the API is the only way to run it. It lists at the same rate as Claude Sonnet 5, takes images as input, and carries a context window of just over a million tokens, five times Sonnet 5's.
In a video pipeline, a language model's job is usually the script: the hook, the beats, the on-screen captions, and the call to action that every later node narrates, renders, and cuts. That is the job we tested it on. This post covers the price per million tokens and per script, the same prompt on Claude Sonnet 5 and Claude Haiku 5.5, where the script went wrong, and how to run the model as the writing step of a pipeline. Every price is the listed rate as of October 2026, and every test number is what the provider billed and what the clock said.
The rate card
| Kimi K3.1 Pro | Claude Sonnet 5 | Claude Haiku 5.5 | |
|---|---|---|---|
| Input, per million tokens | $2.00 | $2.00 | $0.10 |
| Output, per million tokens | $10.00 | $10.00 | $0.50 |
| Context window | 1,048,576 tokens | 200,000 tokens | 1,000,000 tokens |
| Image input | Yes | Yes | Yes |
| Weights | Hosted only | Hosted only | Hosted only |
Two things stand out. The price is Sonnet 5's price to the cent, so this is not a budget model. And the context window is the biggest in its price band: a request can carry hours of transcript or every past script a channel has published for about two dollars of input. Our probe billed exactly at the list rate, so there is no hidden request fee.
The test: one prompt, three models
Every language model that joins the Text Generation node gets the same probe before it is listed: the node's own request with the node's defaults (temperature 0.7, a 1,024 token cap, reasoning off), the system prompt "You write short-form video scripts", and this user prompt:
Write a 30-second vertical video script about why octopuses have three hearts. Use exactly these sections: HOOK (one line), BEATS (4 numbered lines, each under 15 words), ON-SCREEN TEXT (4 captions, each under 6 words), CTA (one line). No emojis, no markdown headers.
We ran it on Kimi K3.1 Pro, on Claude Sonnet 5, the node's default model, and on Claude Haiku 5.5, all on the same morning.
| Kimi K3.1 Pro | Claude Sonnet 5 | Claude Haiku 5.5 | |
|---|---|---|---|
| Prompt tokens | 105 | 122 | 123 |
| Reply tokens | 134 | 218 | 216 |
| Billed by the provider | $0.00155 | $0.0024 | $0.00012 |
| Wall time | 5.1 s | 3.5 s | 2.1 s |
| Finish | Stopped on its own, all four sections | Stopped on its own, all four sections | Stopped on its own, all four sections |
Kimi K3.1 Pro and Sonnet 5 share a list price, so the gap in the billed column is entirely length: 134 reply tokens against 218. It was also the slowest of the three by a clear margin. None of the models spent reasoning tokens or hit the token cap.
What Kimi K3.1 Pro wrote
The reply, unedited:
HOOK: Octopuses have three hearts, and two of them literally stop when they swim.
BEATS:
- Two hearts pump blood through the gills to grab oxygen.
- The third heart pushes that blood everywhere else.
- Blue copper-based blood helps them thrive in cold seas.
- Swimming stops the main heart, so crawling saves energy.
ON-SCREEN TEXT: Three hearts. One weird body. Two hearts serve the gills. One heart serves everything else. Swimming? Exhausting. Crawling? Efficient.
CTA: Follow for more strange animal biology in 30 seconds.
The format held. Four beats, the longest ten words against a limit of fifteen. Four captions, the longest five words against a limit of six. No emojis, no headers. The one difference from the Claude replies is cosmetic: Kimi put a colon after each label and ran the hook and the CTA onto the label's line. A downstream node splitting on the label text handles either shape; a parser that expects a line break after the label will not.
The writing is the tightest of the three. About 58 words of narration against roughly 65 from the Claude models, which leaves room in a 30 second cut for a pause on the hook, and the fourth caption reads like something a Short actually puts on screen.
The fact all three got wrong
The hook says two of the octopus's three hearts stop when it swims. That is backwards. The standard account is that the two branchial hearts keep pushing blood through the gills, and the single systemic heart, the one that serves the body, is what stops during swimming, which is why octopuses prefer to crawl. Kimi's own fourth beat has it right: "swimming stops the main heart". Its hook contradicts its own script.
Sonnet 5 and Haiku 5.5 made the same mistake in their hooks on the same prompt. Three models from two companies, three identical errors, each contradicted later in its own reply. The lesson is the one we drew yesterday for Haiku 5.5: this is an unchecked-fact problem, not a model-quality problem, and paying more for the writer does not fix it.
The pipeline fix is cheap. Wire a second Text Generation node after the writer with a system prompt that checks the hook against the beats and answers with a corrected hook, then feed that into the text to speech step. The checker can be a different model; a Kimi K3.1 Pro script checked by Haiku 5.5 costs about a tenth of a cent more than the script alone. The faceless script guide covers the system prompt rules that keep scripts on the rails in the first place.
The same model on an image
The Text Generation node takes a Media input, so the second probe wired in a product photo of a folded navy linen shirt on an oak table and asked for a description. Kimi K3.1 Pro came back in 9.2 seconds for $0.0049, almost all of it the 2,159 prompt tokens the image occupied, and the description was accurate and compact: the pointed collar, the woven texture, the warm directional light, the shallow depth of field, the wood surface. Sixty-two tokens of reply for a 2,000 token image.
That is the shape of a judge: describe an image, check a generated clip's frame against its prompt, or score a batch of frames against the brief. At this price the image side of the call is about half a cent, fine once per clip and worth a thought once per frame. The OpenRouter listing says the model also accepts video input as of October 2026; our probe sent an image.
Where the script cost sits in a Short
A script on Kimi K3.1 Pro cost $0.00155. On Claude Sonnet 5 it cost $0.0024, and on Claude Haiku 5.5 it cost $0.00012. The Story Video starter in chat is priced at $0.78, and the median charged cost of one video generation on Treza is $1.06. Against either number, all three script prices are rounding error. A channel writing one script a day for a year spends about $0.57 on Kimi K3.1 Pro, about $0.88 on Sonnet 5, and about $0.04 on Haiku 5.5.
So the per-script price does not pick this model, and neither does latency, where it was the slowest of the three. What picks it is the input. A pipeline that hands the writer a two-hour transcript, a season of past scripts to avoid repeating, or a stack of research pages wants a model that reads all of it in one request at a normal price, and a million-token window at $2 per million is that. A Text Generation node with a Transcribe node's output or a fetched page wired in is where Kimi K3.1 Pro earns its place in the picker.
Sonnet 5 stays the node's default because it carries the production record: in September 2026, 1,100 of the 1,173 Text Generation nodes that ran on Treza used it. Kimi K3.1 Pro is the model to try when the prompt is long, the output needs to be tight, or you want a second opinion from a different lab at the same price.
Running Kimi K3.1 Pro as the script step
On a Treza AI video pipeline, the script is the first node in the chain. Set the Text Generation node's model to Kimi K3.1 Pro, put the channel's angle and voice in its system prompt, and turn on Avoid recent repeats so a scheduled channel does not rewrite last Tuesday's video; the node remembers up to 50 previous runs. Set a fallback model so a call that fails after retries moves on rather than failing the run.
From there the script feeds the rest of the line. A text to speech node reads it in one of 82 built-in voices across four models, with 14 languages on the multilingual model, the setup the AI voiceover generator page walks through. Each scene renders as its own clip on the video model you pick, the Sequence node stitches up to 12 shots against the narration, Whisper transcribes the result with word-level timestamps, captions are burned in, and a YouTube or TikTok node publishes it with metadata written from the transcript. The faceless video generator page covers that build, and the script to video page covers the case where the script already exists and the model's job is to split it into scenes.
Publish the pipeline and it becomes a versioned endpoint. Your application calls a typed /invoke route with JSON in and JSON out, or points an existing OpenAI SDK at the OpenAI-compatible route, and our hosted MCP server exposes the same pipeline as a tool an agent in Claude or ChatGPT can call. When the next model at this price writes a tighter script, you change the dropdown and republish, and the endpoint does not change. The node's language models sit alongside the video and image models in the model library.
What you actually pay on Treza
A Text Generation node through a Treza pipeline is metered from a prepaid credit balance. One credit equals $0.01, packs start at $5, credits never expire, and a failed generation charges nothing. Before a run starts, the platform estimates the cost of the graph and refuses one your balance cannot cover. A text node is quoted at a token budget drawn from production history and settles at what the provider billed, with the difference returned to the balance.
For a finished video costed end to end, see the AI video cost breakdown. To try the model on your own brief, open the AI video generator, set the script node to Kimi K3.1 Pro, and run one.
Frequently Asked Questions
How much does Kimi K3.1 Pro cost per million tokens?
$2.00 per million input tokens and $10.00 per million output tokens on the endpoint we run, as of October 2026, with no per-request fee. That is the same listed rate as Claude Sonnet 5 and twenty times Claude Haiku 5.5's. Our probe was billed exactly at the list rate.
How much does a video script cost on Kimi K3.1 Pro?
Our 30-second Short script, a 105 token prompt into a 134 token reply, was billed $0.00155 by the provider and took 5.1 seconds. The same prompt on Claude Sonnet 5 was billed $0.0024 and took 3.5 seconds, and on Claude Haiku 5.5 it was billed $0.00012 and took 2.1 seconds. On Treza the charge comes from a prepaid credit balance, and a failed call charges nothing.
Is Kimi K3.1 Pro good enough to write video scripts?
On our probe it held every structural rule in the prompt: four beats under fifteen words, four captions under six, labelled sections, no emojis, a 58 word read, with the tightest captions of the three models tested. It also put a wrong fact in the hook that its own fourth beat contradicted, the same error Claude Sonnet 5 and Claude Haiku 5.5 made, so put a fact check in the graph whichever model writes.
How does Kimi K3.1 Pro compare with Claude Sonnet 5 for scripts?
Same list price, a reply about 40 percent shorter and so about 35 percent cheaper per script, and about a second and a half slower on our test, with the same four sections and the same factual error. Sonnet 5 remains the node's default because it carries the production record. Kimi K3.1 Pro's advantage is its context window, 1,048,576 tokens against Sonnet 5's 200,000, at the same input price, which is what makes it the pick for a node that reads a long transcript or a season of past scripts in one request.
Can Kimi K3.1 Pro read images?
Yes. Wired into the Text Generation node's Media input, it described a product photo accurately in 9.2 seconds for $0.0049, the image accounting for about 2,150 of the prompt tokens. The OpenRouter listing also says it accepts video input as of October 2026; we tested an image.
How do I use Kimi K3.1 Pro in a video pipeline through an API?
On a Treza pipeline, Kimi K3.1 Pro is a choice in the Text Generation node's model picker. Wire that node to text to speech, video generation, captions and publishing, publish the pipeline, and call it as a versioned HTTP endpoint with a typed /invoke route and an OpenAI-compatible route. A hosted MCP server exposes the same pipeline as a tool for agents. Generations are metered from prepaid credits that start at $5 and never expire.
