Moonshot AI: Kimi K3
Kimi K3 is a reasoning model from Moonshot AI released on Jul 16, 2026. It reads text, images and video and writes text, with a 1.05M token context window and up to 1.05M tokens per reply.
Show moreShow less
It costs $3 per million input tokens and $15 per million output tokens. Cached input bills at $0.30, 90% less than fresh input.
It supports tool calling, structured JSON output and prompt caching, and its weights are open, so you can host it yourself.
72 providers serve it, including SiliconFlow, Deep Infra and Alibaba. The lowest input price is $2.70 per million tokens at SiliconFlow.
On the Artificial Analysis Intelligence Index it scores 43.6, which places it at number 11 among the 43 models I track with independent scores.
- Modalities
- In / out price
- $3 / $15 per 1M
- Context
- 1.05M
- Max output
- 1.05M
- Released
- Jul 16, 2026
- Knowledge cutoff
- n/a
- Tool calling
- Structured output
- Reasoning
- Prompt caching
- Open weights
- Free to use
Kimi K3 pricing
USD, as listed on Oct 15, 2026. Most bills are input heavy, so the cached input rate matters as much as the headline price.
| Input | $3 / 1M tokens | |
|---|---|---|
| Output | $15 / 1M tokens | |
| Cached input (read) | $0.30 / 1M tokens | 90% off input |
| Cache write | $3 / 1M tokens |
Kimi K3 providers
72 providers serve it, including SiliconFlow, Deep Infra and Alibaba. The lowest input price is $2.70 per million tokens at SiliconFlow. Each host sets its own price and limits, so the cheapest one is not always the one with the longest context.
| Provider | Input /M | Output /M | Cache read /M | Context | Max output |
|---|---|---|---|---|---|
| Moonshot AI First party | $3 | $15 | $0.30 | 1.05M | 1.05M |
| SiliconFlow | $2.70 | $13.50 | $0.27 | 1.05M | 262K |
| Deep Infra | $2.85 | $14.25 | $0.285 | 1.05M | 131K |
| Alibaba | $3 | $15 | $0.30 | 1.05M | 1.05M |
| Baseten | $3 | $15 | – | 1.05M | 262K |
| Fireworks AI | $3 | $15 | $0.30 | 1.05M | 131K |
| Hugging Face | $3 | $15 | – | 1M | 131K |
| Nebius Token Factory | $3 | $15 | $3 | 1.05M | 8K |
| Novita AI | $3 | $15 | $0.30 | 1.05M | 1.05M |
| Together AI | $3 | $15 | $0.30 | 1.05M | 131K |
| Vercel AI Gateway | $3 | $15 | $0.30 | 1M | 131K |
| Amazon Bedrock | $3.30 | $16.50 | $0.33 | 1.05M | 128K |
Showing 12 of 72 providers: the lab first, then the major clouds, then the cheapest resellers.
What Kimi K3 costs per month
Pick a workload close to yours or type your own volumes.
Independent benchmarks
Measured by third parties, not reported by Moonshot AI. Each links to its source and shows the date I checked it.
A composite of 10 independent evaluations (agents 30%, coding 20%, scientific reasoning 20%, general 30%), run by Artificial Analysis on each model, version 4.3.2.
Artificial Analysis, checked Oct 10, 2026
Human preference: people compare two anonymous answers and pick the better one. Ratings from the Text Arena overall board, style control on.
Arena (formerly LMArena), checked Oct 10, 2026
Abstract reasoning on visual puzzles the model has not seen, semi private set. Hard to memorise, so it tracks reasoning more than knowledge.
ARC Prize, checked Oct 10, 2026
Expert level questions across many fields, no tools allowed. The Diamond set is 1,000 questions. Few models are on it yet.
Scale Labs, checked Oct 10, 2026
Median output tokens per second measured by Artificial Analysis on the provider API, for the same model setting as its index score.
Artificial Analysis, checked Oct 10, 2026
Benchmarks Moonshot AI reports (11)
- gpqa-diamond (GPQA Diamond, reasoning effort max, temperature 1.0, top-p 0.95)93.5%
- hle-no-tools (HLE-Full, without tools, reasoning effort max)43.5%
- hle-with-tools (HLE-Full, with general tools, reasoning effort max)56%
- terminal-bench-2.1 (Terminal-Bench 2.1, Kimi Code harness)88.3%
- deepswe-v1.1 (DeepSWE v1.1 tasks, Kimi Code harness)67.5%
- browsecomp (BrowseComp, context compaction triggered at 300K tokens)91.2%
- osworld-verified (OSWorld-Verified)84.8%
- mmmu-pro (MMMU-Pro, without tools)81.6%
- toolathlon-verified (Toolathlon-Verified)76.5%
- mcp-atlas (MCP-Atlas, 500-task public subset, 100-turn limit, Gemini 3.1 Pro judge)84.2%
- automationbench (AutomationBench, 600-task public subset)30.8%
Labs pick their own settings and harnesses, so compare these only within one lab.
Supported request parameters
What you can set when you call Kimi K3 through an OpenAI compatible API.
- Tools
tools - Structured outputs
structured_outputs - Reasoning
reasoning
Alternatives to Kimi K3
Models that take the same inputs and call tools, closest in price, cheaper ones first.
- Index 29.7
Reads text, images, files, audio and video. Tool calling, structured output, reasoning. Served by 31 providers.
by googleFeb 18, 20261.05M context$2/M input tokens$12/M output tokens31 providers
- 1M context
Reads text, images and video. Tool calling, structured output, reasoning. Served by 4 providers.
by qwenSep 23, 2026$4/M input tokens$12/M output tokens4 providers
- 1.05M context
Reads text, images, audio and video. Tool calling, structured output, reasoning. Served by 8 providers.
by xiaomiSep 21, 2026$4.35/M input tokens$8.70/M output tokens8 providers
- 1.05M context
Reads text, images, files, audio and video. Tool calling, structured output, reasoning. Served by 30 providers.
by googleFeb 5, 2025$1.25/M input tokens$10/M output tokens30 providers
- Index 25.8
Reads text, images and video. Tool calling, structured output, reasoning. Served by 48 providers.
by moonshotaiJun 12, 2026262K context$1.90/M input tokens$8/M output tokens48 providers
- 1.05M context
Reads text, images, audio and video. Tool calling, structured output, reasoning. Served by 3 providers.
by googleFeb 27, 2026$2/M input tokens$12/M output tokens3 providers
More from Moonshot AI
- 1.05M context
Reads text, images and video. Tool calling, reasoning. Served by 2 providers.
by moonshotaiSep 11, 2026$0/M input tokens$0/M output tokens2 providers
- Index 25.8
Reads text, images and video. Tool calling, structured output, reasoning. Served by 48 providers.
by moonshotaiJun 12, 2026262K context$1.90/M input tokens$8/M output tokens48 providers
- 1.05M context
Reads text, images and video. Tool calling, structured output, reasoning.
by moonshotaiMay 3, 2026$2/M input tokens$10/M output tokens
- 262K context
Reads text, images and video. Tool calling, structured output, reasoning. Served by 59 providers.
by moonshotaiApr 20, 2026$0.95/M input tokens$4/M output tokens59 providers
- 262K context
Reads text and images. Tool calling, structured output, reasoning. Served by 36 providers.
by moonshotaiJan 1, 2026$0.60/M input tokens$3/M output tokens36 providers
- 262K context
Reads text. Tool calling, structured output. Served by 8 providers.
by moonshotaiSep 5, 2025$0.60/M input tokens$2.50/M output tokens8 providers
Kimi K3 questions
How much does the Kimi K3 API cost?
$3 per million input tokens and $15 per million output tokens, with cached input at $0.30. The estimator on this page turns that into a monthly bill for your own workload.
What is the context window of Kimi K3?
1,048,576 tokens of input, and it can write up to 1,048,576 tokens in one reply. As a rough guide, a million tokens is around 750,000 English words.
Which providers serve Kimi K3?
72 providers serve it, including SiliconFlow, Deep Infra and Alibaba. The lowest input price is $2.70 per million tokens at SiliconFlow. Prices differ by host, so the Providers table on this page lists each one with its input and output price.
Does Kimi K3 support tool calling?
Yes. It accepts tool definitions and can return JSON that matches a schema, which is what an agent needs to call your APIs.
Is Kimi K3 open source?
Its weights are published, so you can run it on your own hardware or a cloud of your choice. Read the license before commercial use.
When was Kimi K3 released?
Jul 16, 2026.
How good is Kimi K3 compared with other models?
It scores 43.6 on the Artificial Analysis Intelligence Index and 1488 on the Arena text leaderboard. Benchmarks are a shortlist, not a verdict: test the top two or three candidates on your own tasks before you commit.
Catalog data from models.dev (MIT licensed), refreshed daily, last on Oct 15, 2026. Independent scores from my hand checked rankings, sources linked above. The summary is written from these facts.
Building on Kimi K3?
I test your real tasks on it and the two closest alternatives, then ship the agent on whichever wins on cost and accuracy.
Book Meeting