Kimi K3 AI: Pricing, Benchmarks and Kimi Work

Kimi K3 is Moonshot AI's 2026 flagship model for coding, research, tool use, visual work, and long multi-step tasks. It combines a very large 2.8T-parameter design with a 1M-token context window and can be used through Kimi.com, Kimi Work, Kimi Code, APIs, third-party hosts, or public model weights.

The pricing story needs context. Moonshot's official API rate is $3 per million fresh input tokens and $15 per million output tokens, but cached input is only $0.30 per million. OpenRouter can be much cheaper on some provider routes, and Kimi's app memberships use a separate monthly-credit system.

Kimi Work is also a major part of the K3 story. It turns K3 into a desktop agent that can work with local files, the browser, documents, spreadsheets, slides, plugins, scheduled tasks, and parallel sub-agents.

What Is Kimi K3?

Kimi K3 is Moonshot AI's large multimodal reasoning model for coding, research, tool use, and long multi-step work. It launched in July 2026 and has a 1,048,576-token context window, usually described as a 1M-token context.

The model has 2.8 trillion total parameters. It uses a mixture-of-experts design, which means it does not use all of those parameters for every token. Instead, it routes work through a smaller set of specialist parts of the model.

What Kimi K3 Actually Is

K3 is built for jobs that continue for many steps, such as working through a large codebase, using terminal tools, reading large document collections, checking screenshots, and revising work after tests or feedback.

Moonshot says K3 uses Kimi Delta Attention and Attention Residuals to handle long sequences more efficiently. You do not need to understand those internal methods to use the model. The practical point is that K3 is designed to keep more information available across very long tasks.

What Kimi K3 Includes

K3 supports text and visual input, tool calling, structured output, and long-context reasoning. Exact input support can differ by provider. Some routes expose text and images, while other Kimi surfaces also describe video understanding.

K3 is available through Kimi.com, Kimi Work, Kimi Code, the Kimi API, third-party inference providers, and public model weights for teams that want to host it themselves.

Kimi K3 Pricing at a Glance

Moonshot's official API rate is $0.30 per million cache-hit input tokens, $3.00 per million fresh input tokens, and $15.00 per million output tokens.

The important detail is caching. If a long prompt begins with the same system instructions, tools, or codebase on every turn, a cache hit can reduce that repeated input from $3.00 to $0.30 per million tokens.

App membership and API billing are separate. A monthly Kimi membership does not automatically become API credit.

Official Moonshot API rate card.

Kimi K3 API Pricing, Line by Line

Three items make up most K3 API bills: cached input, fresh input, and output. Output is the most expensive at $15 per million tokens, and K3's reasoning tokens are part of the bill.

Moonshot also lists cache-write charges. A default five-minute cache write is priced at $3 per million tokens, while a one-hour cache write is $6 per million. The longer cache can make sense when the same large document or codebase will be reused over a longer period.

Kimi K3 App Plans and Membership Pricing

According to Lindy's October 5 snapshot of Kimi's US pricing page, the app offered a free Adagio tier plus paid plans at $19, $39, $99, and $199 per month. Annual billing reduced the effective monthly price.

These plans are aimed at people using Kimi's chat, agent, research, document, spreadsheet, slide, Kimi Code, and Kimi Work features. Kimi's official help center says those membership features share one credit pool, so a long Kimi Work task and a Kimi Code session can draw from the same monthly allowance.

Plan names and limits have changed during 2026, so check the live Kimi membership page before buying.

Pricing snapshot reported by Lindy from Kimi's US pricing page on October 5, 2026. Check the live plan page before subscribing.

Which Way to Pay for Kimi K3?

Use a Kimi membership if you mainly want Kimi as a daily chat, research, coding, or desktop-work assistant and prefer a fixed monthly plan.

Use the API if you are building K3 into your own product or internal workflow and want costs tied directly to token usage. A third-party provider can make sense if you already use its billing, routing, regional controls, or deployment tools.

Kimi K3 Pricing on OpenRouter and Other Providers

OpenRouter does not have one fixed K3 price because it routes across many providers. On October 17, 2026, the Kimi K3 page showed a headline price of $0.25 per million input tokens and $8.538 per million output tokens, while Moonshot's own route remained $3 and $15.

That gap is possible because providers use different serving setups, quantization levels, discounts, and routing rules. Treat OpenRouter's headline as a live marketplace price, not the official Moonshot list price.

Together AI and Nebius both list K3 at $3 per million input tokens and $15 per million output tokens for their standard K3 endpoints. Provider availability and prices can change, so check the live endpoint before production.

Why Real Kimi K3 Bills Can Run Higher Than the Rate Card

Price per million tokens is only part of the cost. Long agent tasks can repeatedly send large amounts of previous context, tool output, and reasoning history.

K3 also tends to produce substantial output on difficult tasks. Because output costs $15 per million tokens, a model that writes more can cost more than its simple rate card suggests.

The better measure is cost per accepted result. Include retries, rejected work, human review time, and any fallback call to another model.

The API Math Is the Real Reason to Test It

For a small request with 10,000 input tokens and 2,000 output tokens, Northflank calculates about $0.06 with fresh input. If all input is served from cache, the token cost falls to about $0.033.

For a much larger request with 500,000 input tokens and 20,000 output tokens, the same example is about $1.80 with fresh input and $0.45 if all input is a cache hit.

This is why stable prompt prefixes matter. A clean prompt structure can make a large difference in repeated coding and document-analysis sessions.

Caching can materially reduce repeated long-context input costs.

Kimi K3 Benchmarks: What the Numbers Show

Benchmark results show that K3 is strong at coding, tool use, long tasks, and visual work, but it does not win every test.

Moonshot's launch evaluation reported 67.5 on DeepSWE, 77.8 on Program Bench, 88.3 on Terminal-Bench 2.1, 81.2 on FrontierSWE, and 42.0 on SWE Marathon. These are different tests, so the scores should not be averaged together.

OpenRouter's current page also reports strong results across GPQA Diamond, coding tests, agent tests, and multimodal benchmarks. Because benchmark suites and scoring methods change, always compare scores from the same source and date.

Selected launch coding benchmarks. These tests measure different tasks and should not be combined into one score.

The Benchmarks Say Kimi K3 Is Competitive, Not the Best at Everything

Moonshot itself says K3 still trails the strongest proprietary models in overall performance and user experience. That is useful context because one strong benchmark row does not make a model the best choice for every job.

BenchLM's October 16, 2026 snapshot gave K3 an overall score of 69.9 out of 100 and ranked it 18th of 224 tracked models, while listing multimodal and grounded work as its strongest category. The exact rank will change as new models and benchmark results arrive.

How Good Is Kimi K3?

K3 is most convincing when the task is long, has many steps, and benefits from large context. Examples include repository-wide coding, research across many documents, tool-driven analysis, and multimodal workflows that use screenshots or images.

For short, simple requests, K3 can be more model than you need. A cheaper or faster model may deliver the same practical result.

Kimi-K3 for Long-Running Agents

Nebius describes K3 as a model for demanding reasoning, coding, tool use, and multimodal work across a 1,024K-token context. It highlights long-running agents, research, and coding as core use cases.

An agent is simply software that lets the model plan steps and use tools instead of answering once. K3 is designed to keep working through these multi-step loops for much longer than a normal chat response.

What Is Kimi Work?

Kimi Work is Moonshot's desktop agent for knowledge workers. It is available in the Kimi desktop client on Windows and Apple silicon Macs.

You describe a goal in normal language. Kimi Work can then break the job into steps, use the browser, work with selected local files, call plugins or skills, and create deliverables such as documents, spreadsheets, slide decks, dashboards, and project files.

The word local refers to the agent working with files and applications on your computer. It does not mean the full 2.8T K3 model is running on a laptop.

A simplified view of how Kimi Work turns a goal into a finished deliverable.

How Kimi Work Uses Kimi K3

Kimi Work launched in June 2026 and has used Kimi K3 since version 3.1.0. The current release notes list version 3.2.17 on October 15, 2026.

K3 gives Work the ability to stay with long tasks, use tools repeatedly, and coordinate parallel work. Kimi Work can run a normal Agent or Agent Swarm, where multiple sub-agents work on parts of a larger task.

Kimi Work Features That Matter in Practice

Kimi Work can read and create files in selected local folders, automate browser tasks, run plugins and skills, schedule tasks, manage projects, and show results in dashboards.

Recent releases added a built-in browser, remote control from a phone, Office file editing, file version history, side chats, annotations, and better support for long-running background work.

Kimi Work also has permission controls. You can require approval before sensitive actions such as changing files or running code, or allow more automatic execution when you trust the workflow.

Kimi Work vs. Kimi.com vs. Kimi Code

Kimi.com is the easiest choice for chat, research, files, and general online agent work.

Kimi Work is better when the task needs local files, desktop apps, browser automation, scheduled work, or a long-running job on your computer.

Kimi Code is aimed at developers working in terminals and supported IDEs. It understands codebases, edits files, and runs development commands.

How Do You Use the Kimi K3 API?

The Kimi API uses an OpenAI-compatible interface. The Moonshot base URL is https://api.moonshot.ai/v1 and the model ID is kimi-k3.

A basic request sends messages to the Chat Completions endpoint. K3 also supports tools and structured output. For multi-turn agent sessions, Moonshot recommends passing the complete assistant message back into the next turn so the model keeps the reasoning and tool history it expects.

Can You Self-Host Kimi K3 Today?

Yes. Early launch articles said the weights were still coming, but that changed on July 27, 2026. Northflank's updated article confirms that Moonshot released the public weights and production-ready Kimi Delta Attention support for vLLM.

Self-hosting is possible, but it is not a small deployment. Moonshot's launch guidance recommended at least 64 accelerators for serving the full model. That makes K3 self-hosting a data-center or large-cluster project rather than a normal workstation setup.

Choose How to Run Kimi

Use Kimi.com for quick access. Use Kimi Work when local files and desktop workflows matter. Use Kimi Code for software development. Use the Kimi API when you are building a product. Use a third-party host when you want different routing, regions, or billing. Self-host only when you have the infrastructure and a strong reason to control the model deployment.

Choose the access method based on the job and infrastructure you need.

Kimi K3 Limitations to Know

Moonshot lists three important limitations. First, quality can become unstable if a tool or agent does not preserve the thinking history K3 expects across turns.

Second, K3 can be overly proactive on ambiguous tasks. If you need strict boundaries, give the model clear rules about what it may and may not change.

Third, Moonshot says the overall user experience still has a noticeable gap compared with the strongest proprietary systems. Independent reviews also note that K3 can be slower and more verbose than some alternatives.

Is Kimi K3 Worth the Price?

K3 makes the most sense when you need what is unusual about it: a 1M-token context, strong multi-step coding, visual understanding, open weights, and a desktop agent that can work with local files.

It is harder to justify for short, high-volume tasks where cheaper models can provide similar quality. The right decision depends on cost per finished result, not the headline token rate.

For teams, a controlled trial is usually smarter than changing every workload at once. Test K3 on real tasks, measure success rate, review time, latency, and total cost, then decide where it earns a place.

How to Cut Your Kimi K3 Bill

Keep stable content at the beginning of prompts so context caching can work. Avoid putting timestamps or frequently changing text before a large reusable prefix.

Use a lower reasoning effort when your endpoint supports it and the task does not need the maximum setting. Limit unnecessary output, monitor token usage, and measure cost per accepted result.

For personal use, compare API spend with Kimi's monthly membership. A fixed plan can be cheaper if you use chat, Kimi Work, or Kimi Code every day.

Animated Kimi Work Workflow

The animation below highlights the five main stages of a Kimi Work task, from the initial goal to the finished files.

Animated WebP: goal, plan, run tasks, review, deliver.

Frequently Asked Questions

Final Thoughts

Kimi K3 is not simply a cheaper chat model. Its strongest argument is the combination of very large context, strong multi-step coding and research, public weights, and Kimi Work as a desktop execution layer.

The official $3/$15 API rate is not cheap in absolute terms, but caching can change the economics of repeated long-context work. OpenRouter and other hosts can also offer different rates, so the actual provider matters.

For most teams, the best approach is to test K3 on a few real workflows, measure the total cost per accepted result, and decide where its long-task strengths are worth the price.

Sources

• • OpenRouter: MoonshotAI Kimi K3
  • Lindy: Kimi K3 Pricing
  • Omid Saffari: Kimi K3 Review
  • TrilogyAI: Kimi K3 Is Live
  • Northflank: Kimi K3 Benchmarks, Pricing and Self-Hosting
  • Reddit discussion: Kimi K3 pricing
  • Nebius: Kimi Models Inference
  • BenchLM: Kimi K3
  • Together AI: Kimi K3
  • Kimi official: Kimi K3 Tech Blog
  • Kimi Help Center: Kimi Work Overview
  • Kimi Help Center: Kimi Work Release Notes
  • Kimi Help Center: Product Comparison