Kimi K3 is Moonshot AI’s most capable model, and this Kimi AI review answers one question: is it worth paying for? The Kimi app starts free on the Adagio plan, and the paid plans run from $19 to $199 a month. On the API, developers pay by the token, a chunk of text a few characters long: $3 per million going in and $15 per million coming out. The weights have been free to download since July 27, but they weigh 1.56 TB.
Our take: start free, and pay $19 for Plus when you want Kimi Code or Agent Swarm, which the free plan doesn’t include. If you want cheap API tokens, K3 isn’t the pick: per token it costs more than Claude Sonnet 5.5. And the open weights are for companies with server clusters, not for your laptop.
⚡ TL;DR: Kimi K3 in One Minute
What It Is: Moonshot AI’s Kimi app and API, now running K3, a 2.8T-parameter model whose weights anyone can download.
Best For: People who want Kimi’s agents and Kimi Code for $19 a month, and companies that want to host a frontier-class model on their own servers.
Price: Free (Adagio). Plus $19, Pro $39, Max $99, Ultra $199 a month, or $15 to $159 a month billed yearly. API: $3 in, $15 out per million tokens.
Our Take: Start free. Plus earns its $19 if you will use Kimi Code or Agent Swarm; move up only when Plus runs out of credits.
⚠️ The Catch: Kimi’s pricing and help pages give no credit number for any plan, unused credits don’t roll over, and a paid plan can’t be refunded once active (an upgrade credits the unused days).
📑 Quick Navigation
What is Kimi K3, and what changed since K2.6?
Moonshot launched K3 on July 16, 2026. It is a 2.8T-parameter model that reads images natively and has a 1-million-token context window. Think of the context window as the model’s desk: it is how much text the model can keep in view at once before older pages fall off.
Moonshot’s launch post is frank about where K3 stands. It says K3’s overall performance still trails Claude Fable 5 and GPT 5.6 Sol, and it lists “a noticeable gap in user experience” against both as a known limitation.
Two things changed in the Kimi app, and one for developers:
- The app has a model switcher. Kimi’s help center lists K2.6, K3 and K3 Cluster, and all three draw on credits, including K2.6 in plain chat.
- The weights arrived on time. For developers and companies, this is the big one: Moonshot promised them by July 27, and its Hugging Face repository opened that day.
How much does Kimi cost in USD? Every plan
These are the prices on Kimi’s international pricing page as of October 17, 2026, in US dollars. Every paid plan lists Kimi’s agents for documents, slides, spreadsheets, Deep Research and website deployment; the table shows what changes as you move up.
| Plan | Monthly | Billed yearly | What it adds |
|---|---|---|---|
| Adagio | $0 | $0 | 1 agent task at a time, 2 projects, 500MB storage |
| Plus | $19 | $15 a month ($180) | Kimi Code, Agent Swarm (2 subtasks), 2 tasks at a time, 20 projects, 20GB |
| Pro | $39 | $31 a month ($372) | “2x agent credits”, Goal mode, one-click Kimi Claw, Swarm with 4 subtasks |
| Max | $99 | $79 a month ($948) | “5x agent credits”, 4 tasks at a time, Swarm with 8 subtasks, chats up to 1M tokens, 50GB |
| Ultra | $199 | $159 a month ($1,908) | “10x agent credits”, 4 tasks at a time, Swarm with 8 subtasks, chats up to 1M tokens, 50GB |
The multipliers are the catch. The page promises Pro “2x agent credits” and Ultra “10x”, but neither it nor the help pages we read on October 17 give a credit number for any plan. So you can’t work out in advance how many slide decks or research reports a month buys.
Billing yearly saves $48 a year on Plus ($180 instead of $228) and up to $480 on Ultra. Paying for Moderato or Allegretto? Kimi’s help center lists Andante, Moderato, Allegretto and Allegro as legacy plans (on a page that describes its China plans): subscribers keep their plan, its rules and its auto-renewal, and switching is optional.
Is Kimi’s free plan enough?
For trying Kimi, yes. For regular agent work, probably not. Adagio costs nothing and runs one agent task at a time, with 2 projects and 500MB of project storage. It has no Kimi Code and no Agent Swarm; both start at Plus.
What Kimi’s pages don’t say is how far free goes, or whether Adagio includes K3. The pricing page lists no credit allowance and no models per plan, and the help center says even K2.6 chat uses credits. The practical move: keep the reasoning effort on Standard, which Kimi’s own guide recommends to save credits.
Say you mostly ask quick questions and draft emails. Then K2.6 is the sensible default: Kimi describes it as the faster option for quick Q&A, while K3 is the one that builds editable .pptx, .docx and .xlsx files. If free runs dry, compare it with the other plans in our roundup of free AI tools and their catches.
What $19 a month buys: Kimi Code and Agent Swarm
Plus adds the two features free lacks. Kimi Code is Moonshot’s coding assistant for the terminal and VS Code. Agent Swarm splits one big job across several AI agents (helpers that do the steps on their own) working at once: 2 subtasks on Plus, 4 on Pro and 8 on Max and Ultra.
In the app’s model switcher, the many-agent option is K3 Cluster. Kimi says its agent cluster can coordinate up to 300 subagents and more than 4,000 parallel tool calls, for jobs like large-scale search or batch processing. Kimi Code has its own rolling 5-hour usage limit (no weekly limit on the current plans), and running out of the shared monthly pool stops it too.
The office tools spend from the same pot. Slides, documents, Deep Research, Kimi Code, K3 and K3 Swarm all draw on one shared credit pool, and if one feature uses it up, the others stop too. A week of heavy slide work can leave you short for coding. Our Kimi Slides review covers the deck generator itself.
The fine print on every paid plan
- No rollover. Credits refresh monthly, even on yearly plans, and unused credits are cleared.
- No refunds. Once a plan is activated, Kimi treats it as consumed: no refund or transfer unless your local law says otherwise. Upgrading is the exception: Kimi refunds the unused days of your current plan.
- Kimi Claw can keep billing. Kimi’s help center says a deployed Claw cloud host costs about 0.6% of your credits a day until you delete it (the page lists the China plans).
- Top-ups exist. Run out and you can switch on an Extra Usage pack that bills as you go.
Say you’re a student who builds a few slide decks a week and wants Kimi Code for assignments. Plus covers both features for $19. Whether its credits cover your month is something you find out afterwards: Settings, then Subscription shows your balance and your last 1,000 usage records.
What does Kimi K3 cost on the API?
The API is how developers plug K3 into their own apps and code. On Moonshot’s API pricing page, K3 costs $3 per million input tokens, $15 per million output tokens and $0.30 per million cached input tokens: about 3.2 and 3.75 times K2.6’s input and output rates. A token is a chunk of text, roughly 3-4 English characters. The rate stays flat across the full 1,048,576-token context, and it is billed pay-as-you-go, separately from the Kimi membership.
The bigger cost driver is thinking. K3 always has thinking switched on, and in Simon Willison’s day-one run the reasoning was counted as output: one SVG drawing used 95 input tokens and 16,658 output tokens, 13,241 of them reasoning, for a total of 25 cents. At $15 per million, those output tokens alone account for almost all of that $0.25.
You can turn the thinking down. The API’s reasoning_effort setting takes low, high or max, and the default is max, so a developer who never touches it pays for the most thinking on every call.
Say you use the API: two monthly bills compared
Say you send 30 requests a day for a month, each with 2,000 input tokens and 5,000 output tokens, thinking included. That is 1.8 million input and 4.5 million output tokens a month. Now say a coding tool sends 20 million input tokens a month, 90% of them cache hits (Moonshot’s own figure for coding work), and gets 2 million output tokens back.
| Model (per million: input / output / cache hit / cache write) | A: 30 requests a day | B: coding tool with caching |
|---|---|---|
| Kimi K3 ($3 / $15 / $0.30 / $3) | $72.90 | $41.40 |
| Kimi K2.6 ($0.95 / $4 / $0.16 / no write fee listed) | $19.71 | $12.78 |
| Claude Sonnet 5.5 ($2 / $10 / $0.10 / $2.50) | $48.60 | $26.80 |
The arithmetic for scenario A on K3: 1.8 x $3 + 4.5 x $15 = $72.90, about 3.7 times the K2.6 bill. Scenario B prices the 18 million cached tokens at each model’s cache-hit rate and the other 2 million at its cache-write rate (K2.6 lists no write fee, so its normal input rate). The table assumes every model writes the same number of tokens. Against other models that may flatter K3: Artificial Analysis, an independent firm that scores AI models, found it “notably slow and somewhat verbose”. Against K2.6 it cuts the other way, because the same firm measured K3 using 21% fewer output tokens than K2.6, so the K2.6 gap is probably overstated.
📊 Monthly API Bill for the Same Work, at List Prices (USD)
💡 Key Takeaway: If you build on K3, set reasoning_effort below max and keep your prompts cacheable. Even then, at list prices it costs more per token than Claude Sonnet 5.5, and K2.6 at $0.95 and $4 is still on Moonshot’s price list for jobs that don’t need K3.
Can you download Kimi K3 and run it yourself?
You can download it. Running it is another matter. The full weights went up on Hugging Face on July 27, 2026, under the Kimi K3 License. The files total 1.56 TB, and Moonshot recommends serving K3 on clusters of 64 or more accelerators (AI chips).
Say you use GPUs with 80 GB of memory each. The files come to 1,560 GB, and 1,560 / 80 = 19.5, so you need 20 cards just to hold them, before any memory for the conversation itself. For scale, Simon Willison runs open models on a 128GB MacBook Pro when they fit; K3’s files are more than 12 times that.
So the weights are for companies and research labs with their own clusters, cloud providers, and teams that need a frontier-class model on hardware they control. Moonshot recommends vLLM, SGLang or TokenSpeed to serve it. Everyone else uses K3 through Kimi’s app, its API or a hosting provider. If you want an open model sized for a home machine, our Gemma review covers that end of the market.
What the Kimi K3 License allows
- Most uses are free. Anyone may use, copy, modify, sell, run, deploy and fine-tune (further train) the weights at no charge.
- Big API hosts need a contract. If a company or its affiliates run a “Model as a Service” business (selling API access to models) and the group’s revenue passes $20 million over any 12 months, it needs a separate agreement with Moonshot before any commercial use of K3.
- Huge products must credit it. A product with over 100 million monthly active users, or over $20 million in monthly revenue, must show “Kimi K3” prominently.
- Internal use is exempt. None of those conditions apply when neither the model nor its outputs reach third parties.
🔍 REALITY CHECK
The claim: “It is the world’s first open 3T-class model”
The fine print: Open, yes: the weights are public and the license is free for most uses. But open doesn’t mean it runs at home. The files are 1.56 TB, Moonshot recommends 64 or more accelerators, and big API hosts need a separate agreement.
Verdict: Treat K3 as open for companies and researchers. For one person, the app or the API is the realistic way in.
Our July 2026 video came out before the weights did, so its lead point, that you couldn’t download K3, no longer holds: Moonshot published them on July 27. Its pricing points still match Moonshot’s API page as of October 17. One comparison has moved: in July those rates put K3 level with Claude’s Sonnet models, and Sonnet 5.5 lists at $2 and $10.
Is Kimi safe for your data?
Read two lines on Kimi’s data usage page before you paste anything sensitive. It says user inputs and responses may be used for model training once de-identified, and that user data is stored on servers in mainland China. Kimi’s privacy policy is issued by Beijing Moonshot Technology Co., Ltd.
You can opt out of training by emailing the support address on that page. Registration takes 5-7 working days and covers your whole account, not single chats. On the business plan, Kimi Business, enterprise data stays out of training.
Our rule: keep client work, health records and anything under a confidentiality agreement out of the consumer app. If you need K3 for that kind of work, run the open weights on infrastructure you control. That is what the open weights are for.
Kimi K3 reviews: what independent reviewers found
If you came looking for Kimi K2 Thinking user reviews and feedback, K3 is the model to judge: Kimi’s help center lists K2.6, K3 and K3 Cluster as the app’s models. We found no large sample of user reviews, so here is what two independent reviewers report. Artificial Analysis calls K3 “particularly expensive when comparing to other open weight models of similar size”. Simon Willison found its vision works well and its pricing steep: his single drawing cost 25 cents.
Moonshot’s own limitations list is worth reading before you hand K3 an agent job. It warns that K3 “may make unexpected decisions on the user’s behalf” and advises against switching to K3 in the middle of a session. Its own advice is to set explicit boundaries in your instructions and to start a fresh chat rather than switch models midway.
Kimi AI FAQ: free plan, USD prices and safety
Is Kimi AI still free?
Yes. The Adagio plan costs $0 and runs one agent task at a time with 500MB of project storage, but it has no Kimi Code or Agent Swarm. Kimi’s pricing and help pages (read on October 17) list no credit allowance for it and don’t say whether it includes K3.
How much is Kimi AI in USD?
Plus is $19 a month, Pro $39, Max $99 and Ultra $199, or $15, $31, $79 and $159 a month billed yearly. The K3 API is billed separately, at $3 in and $15 out per million tokens.
Is Kimi K2 free?
Not without limits. Kimi’s help center says K2.6 in chat draws on membership credits like every other feature, and its pages don’t say how many credits the free plan gets. On the API, K2.6 costs $0.95 in and $4 out per million tokens.
Kimi Moderato vs Claude Pro: which is better value?
Moderato is now a legacy plan; Kimi’s current $19 plan is Plus, against $20 a month for Claude Pro and $20 for ChatGPT Plus. The $1 gap matters less than the terms: Kimi’s pages give no credit numbers, its help center says user data is stored in mainland China, and an active plan can’t be refunded.
Is Kimi fake or a scam?
No. Kimi is run by Moonshot AI (its privacy policy is issued by Beijing Moonshot Technology Co., Ltd.), the lab that publishes the K3 weights on Hugging Face. The risks are ordinary ones: a plan can’t be refunded once activated, and your data is stored in mainland China.
What’s the best Kimi provider for K3?
Moonshot’s own API sets the reference price: $3 in and $15 out per million tokens, with cache hits at $0.30. To host it yourself, Moonshot recommends vLLM, SGLang or TokenSpeed. We didn’t compare third-party hosts’ prices for this update.
Kimi AI review verdict: should you pay for Kimi?
Kimi K3 is a strong model from a vendor that publishes its own limits, sold through a pricing page that leaves out the one number you need. Our call, by reader:
- Curious, light user: stay on the free Adagio plan and see whether it covers your month.
- You want agents and Kimi Code: Plus at $19 a month ($15 billed yearly) is the plan. Move up only after Plus has run out of credits a couple of times.
- Developer on the API: price your workload first. K3 costs more per token than Claude Sonnet 5.5, and K2.6 at $0.95 and $4 still covers lighter jobs. For coding on a budget, compare a flat plan such as the GLM Coding Plan.
- Company with GPUs and privacy needs: the open weights are the reason to pick K3. Read the license’s revenue clauses first.
✅ What We Liked
- ✓ Weights downloadable, free for most uses under the Kimi K3 License
- ✓ Moonshot publishes its own limitations, including where K3 trails rivals
- ✓ Plus at $19 includes Kimi Code and Agent Swarm
- ✓ One flat API rate across the full 1M-token context
- ✓ Yearly billing saves up to $480
❌ What Fell Short
- ✗ No credit number for any plan on Kimi’s pricing or help pages
- ✗ No refunds once active (upgrades excepted), no credit rollover
- ✗ API costs more per token than Claude Sonnet 5.5
- ✗ User data stored in mainland China, per Kimi’s help center
- ✗ 1.56 TB of weights need data-center hardware
Try Kimi free at kimi.com for a week before you pay, and note which features you reach for. For another open model you can download, see DeepSeek vs ChatGPT.
Founder of AI Tool Analysis. Reads every pricing page, help article and footnote so you don’t have to. Covering AI tools since 2025. See how we research →
Related Reading
- Kimi AI Slides Review
- DeepSeek vs ChatGPT
- Claude AI Review
- ChatGPT Review
- Gemma Review: an Open Model for Home Hardware
- GLM Coding Plan Review
- Claude Code vs Cursor
- Free AI Tools: What Each Free Plan Really Gives You
Stay Ahead of the AI Curve
Don’t miss the next big update. Get AI Weekly delivered every Thursday at 9 AM EST with honest reviews, price drops, and features you can actually use.
- ✅ Sourced Reviews: Every price and limit checked against the vendor’s own pages, not rewritten from press releases
- ✅ Price Tracking: Know when tools drop prices or add free tiers
- ✅ Feature Launches: Major updates covered within days
- ✅ Comparison Updates: As the market shifts, we update our verdicts
- ✅ No Hype: Just the AI news that actually matters for your work
Free, unsubscribe anytime.
Last Updated: October 17, 2026
Facts checked: October 17, 2026 (prices, plans and limits against Moonshot’s own pricing, API, help and Hugging Face pages)
Next Review Update: November 2026
Have a tool you want us to review? Suggest it here | Questions? Contact us