Kimi K3.1 Pro Guide: A Near-Frontier AI at Half the Price of Claude Opus — and the 3D Product Play

Moonshot AI just hired you a senior engineer who also sculpts. Kimi K3.1 Pro launched on October 23, 2026 — Moonshot’s most capable model to date, in what the company calls its “3T class” (roughly 3 trillion parameters, the knobs the AI tunes while learning). Think of a model as a new employee: the parameter count is roughly how much it has studied, and this one has studied a library the size of a small country. It writes full software projects, builds 3D models from a sentence, and works through long financial documents — and it costs about half what Anthropic charges for its flagship.

Three words, explained once (then we move on)

• Open-weight: the model’s “weights” are the files that make it what it is — like a recipe plus a trained chef’s instincts. “Open-weight” means Moonshot will publish those files so anyone can download and run the model on their own computers. K3.1 Pro’s weights are promised by November 4, 2026 — they are not public yet, so today you can only use the hosted version.
• Mixture-of-experts (MoE): instead of one giant brain thinking about everything, the model is a panel of specialists and only a few speak up for each word. K3.1 Pro has about 2.8 trillion parameters in total but only about 104 billion “active” per word. What this means for you: it can be very capable without being ruinously expensive to run.
• Context window: how much the model can hold in mind at once. K3.1 Pro’s is 1 million tokens — roughly 750,000 English words, or a shelf of about 15 long novels. What this means for you: you can hand it an entire codebase, a year of financial statements, or a 200-page manual in one go and it will still remember page one.

Why should you care?

Two reasons, and they compound. First, the price gap is real: at Moonshot’s list prices, K3.1 Pro is $2 per million input tokens and $10 per million output tokens — half of Claude Opus 5.5’s $4/$20, and the same as Claude Sonnet 5.5. If you sell AI work by the job, your raw material just got cheaper while staying near the top of the quality ladder. Second, the weights are coming: when a near-frontier model becomes downloadable on November 4, anyone with a good GPU setup can run a top-tier AI without paying a lab per word — and the services built around it (setup, fine-tuning, hosting) become sellable skills. What this means for you: this week you can use it cheaply; next month you may be able to run it yourself.

Price and free quota (all prices vendor-stated, not independently verified)

Item	Price	What this means for you
Kimi K3.1 Pro	$2 input / $10 output per million tokens	A typical research job (20,000 tokens in, 10,000 out) costs about $0.14. The model is not your cost bottleneck — your time is.
Cached input	$0.20 per million tokens	Reusing the same instructions is 50× cheaper than fresh input. What this means: build one good prompt and reuse it, and repeat jobs get nearly free.
Kimi K3.1 Pro Highspeed	$4 input / $20 output per million tokens, same $0.20 cache	Runs 2–3× faster (Moonshot’s claim, not independently verified). What this means: pay double only when a human is waiting on the answer; overnight batch jobs should use the standard model.
Kimi K3 (July model, for comparison)	$3 input / $15 output per million tokens	K3.1 Pro is one-third cheaper than its own predecessor. What this means: if you built anything on K3, switching is a free upgrade in both price and quality.

Free tier, honestly stated: Moonshot does not publish a free API tier for K3.1 Pro. The web app at kimi.ai is the no-cost way to try it, and API access unlocks after a top-up (guides derived from Moonshot’s docs report a $1 minimum top-up). Enterprise access is off by default — an admin must enable it. What this means for you: budget $1 to start, not $0 — and test on the web app before you spend anything.

Hands-on tutorial: your first result in 10 minutes (no code)

Step 1 — Open kimi.ai and sign in. Go to kimi.ai in your browser and create an account (email sign-up). You will land in a chat window, like any AI chat app.

Step 2 — Pick the model. In the model menu, select Kimi K3.1 Pro. If you only see other models, update the app or check back — Moonshot says it rolled out across kimi.ai, the mobile apps, Kimi Work, and Kimi Code on launch day.

Step 3 — Try the 3D maker (the fastest wow). This is the feature that makes K3.1 Pro different from a plain chatbot. Copy and paste this prompt:

Design a small ceramic coffee mug with a matte sage-green glaze, a chunky handle, and a tiny embossed mountain on one side. Make it as a 3D model I can preview, edit, and download for 3D printing.

Step 4 — What you should see. K3.1 Pro’s 3D Make feature (live on kimi.ai and the Kimi app) turns the description into a 3D model you can rotate on screen, edit, download for 3D printing, or export as a video or Live Photo. If the first shape looks off, reply with one correction — “make the handle thicker” — and watch it revise rather than starting over.

Step 5 — Try a work task. Paste this second prompt to see the long-context skill:

I run a small online shop. Write me a one-page product description for this mug, plus three Instagram captions and a $49 launch offer. Keep it friendly, no jargon.

You now have a sellable asset (a 3D model plus copy) made in under 10 minutes, for free on the web app.

For sellers: your first API call (copy-paste)

If you want to build a service on K3.1 Pro rather than chat with it, use the Kimi API. Steps: (1) sign up at platform.kimi.ai; (2) top up (reported $1 minimum) and create an API key — treat it like a password; (3) run this from a terminal, replacing YOUR_KEY:

curl https://api.moonshot.ai/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3.1-pro",
    "messages": [{"role": "user", "content": "Write a 100-word product description for a handmade ceramic mug."}]
  }'

Notes from the docs: thinking is always on for K3.1 Pro (effort levels low/high/max, default high on the coding surface), and keys issued on platform.kimi.ai only work with the api.moonshot.ai host — do not mix hosts or you will get a 401 error. Vision input does not accept public image URLs; use base64 or uploaded file IDs.

Make money with it: the $199 product-configurator service

The customer: small Shopify/Etsy sellers and local furniture or gift shops whose products come in variants (colors, sizes, engravings) but whose listings show one flat photo.

What you sell: a “3D product page” — a rotatable 3D model of their bestseller (built with K3.1 Pro’s 3D generation and parametric modeling), plus the product copy, for $199 per product, delivered in 48 hours.

Why the math works: your AI cost per product is under $1 at list prices (a few long prompts plus revisions). Your real cost is the hour you spend picking the best render, cleaning the description, and packaging the files. At $199, even after a revision round, you keep over $150 per product for about an hour of work.

7-day validation plan: Day 1–2, make 3 free samples from your own or a friend’s products. Day 3, post the rotating models in one Etsy seller group and one local business group with the line “I turn one product photo into a 3D listing — $199, 48 hours.” Day 4–7, take the first 2 paying orders; if nobody bites, drop the price to $99 for the next 3 and ask each prospect what stopped them. Two paying customers in 7 days means the offer works; zero replies after 20 conversations means the pitch, not the model, needs fixing.

Pitfalls and limits (read before you promise a client anything)

• Weights are not out yet. Do not sell “we run it on your own servers” until November 4 — and even then, check the license first: no K3.1 Pro license had been published at launch, so commercial-use terms are unconfirmed.
• It deliberately lost some hacking skill. Moonshot removed cyber-related data from training so the open release would not extend the frontier of hacking capability. What this means for you: do not pitch it as a security-testing tool — that is the one capability Moonshot intentionally trimmed.
• Reasoning tokens are the bill. K3.1 Pro always reasons before answering, and reasoning tokens count as output. One independent test found 85–90% of tokens in a short answer were reasoning nobody reads. What this means: keep prompts tight and prefer the standard model over Highspeed for batch work.
• Data residency is unclear. Moonshot is a Beijing company and publishes no per-region processing map for its API; Zero Data Retention is available to API customers, but treat hosted usage as processing outside the US/EU unless you have it in writing. Do not run client medical or legal documents through it without checking.
• Benchmarks are vendor claims. Every score above (Terminal-Bench 65.7%, BenchCAD 92.7%, OfficeQA Pro v2 64.4) comes from Moonshot’s own tests unless marked independent. The independent Artificial Analysis score (55, fourth overall) is the number to quote to skeptics.