Kimi K3 API Pricing Guide 2026: $3 Input, 1M Context, Vision, and Caching

Moonshot AI released Kimi K3 on July 16, 2026. It is an always-reasoning mixture-of-experts model built for long-horizon coding, research, tool use, and multimodal agents.

The global API costs $3.00 per million cache-miss input tokens, $0.30 per million cached input tokens, and $15.00 per million output tokens. The China API lists the same model at ¥20/¥2/¥100 per million tokens. This guide keeps those regions separate instead of converting currencies.

Kimi K3 Price and Limits

Item	Kimi K3
API model ID	kimi-k3
Global input	$3.00 / 1M tokens
Global cached input	$0.30 / 1M tokens
Global output	$15.00 / 1M tokens
China input / cached / output	¥20 / ¥2 / ¥100 per 1M
Context window	1,048,576 tokens, input plus output
Default maximum completion	131,072 tokens
Maximum configurable completion	1,048,576 tokens
Input	Text, image, video
Tool and output support	Function calling, JSON mode, strict JSON Schema
Reasoning	Always enabled; reasoning_effort: "max"

K3 uses flat token prices across its full context window. Moonshot does not publish a separate long-context surcharge.

OpenAI-Compatible API Setup

Kimi’s global Chat Completions endpoint is OpenAI-compatible. Install the OpenAI SDK and change the base URL:

npm install openai

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.MOONSHOT_API_KEY,
  baseURL: "https://api.moonshot.ai/v1",
});

const response = await client.chat.completions.create({
  model: "kimi-k3",
  messages: [
    {
      role: "user",
      content: "Review this migration plan and return the three highest-risk steps.",
    },
  ],
  max_completion_tokens: 4096,
});

console.log(response.choices[0].message);

Use the endpoint and key shown in your Kimi platform account. Do not mix global and China API keys or base URLs.

Automatic Context Caching

Kimi API automatically detects reusable prompt prefixes. There is no cache ID to create and no TTL to configure. Keep the system prompt, tool definitions, repository summary, and other stable content at the beginning of each request so repeated calls can match the same prefix.

Moonshot reports cache-hit rates above 90% in its own coding workloads. Treat that as a vendor result, then measure your production hit rate from actual usage.

For a task with 2M input tokens and 500K output tokens:

• No cache hits: 2 × $3 + 0.5 × $15 = $13.50
• 90% of input cached: 1.8 × $0.30 + 0.2 × $3 + 0.5 × $15 = $8.64

The output bill remains the largest part of this example. Capping unnecessary reasoning and final-answer length still matters, even with strong input caching.

Architecture and Capability Notes

K3 has 2.8 trillion total parameters and activates 16 of 896 experts per token. Moonshot describes it as the first open model in the three-trillion-parameter class. Its architecture combines Kimi Delta Attention, Attention Residuals, and Stable LatentMoE.

The practical API features matter more than the raw parameter count:

• Native image and video understanding
• Function calling and tool_choice
• Strict structured output through JSON Schema
• JSON mode and Partial Mode
• Dynamic tool loading for large tool catalogs
• Automatic context caching
• A 1M-token context window

K3 currently always reasons at maximum effort through the public API. The official quickstart lists only max; do not assume lower API reasoning levels are available just because another Kimi product exposes model options.

In a multi-turn conversation or tool loop, send the complete prior assistant message back on the next request, including reasoning_content, content, and tool_calls. Moonshot warns that dropping the reasoning history or switching into K3 midway through a reasoning chain can make output unstable.

Open-Weight Status

Moonshot announced that full K3 weights will be available by July 27, 2026. As of this article’s July 19 verification date, that release date is still in the future.

This distinction matters for infrastructure planning: K3 is announced as an open model, but teams should not write a self-hosting plan as though the full checkpoint is already downloadable. Recheck the official K3 launch page after July 27.

Is K3 the Best Coding Model?

K3 is a serious long-context coding and agent candidate, but Moonshot’s own launch post says its overall performance still trails the strongest proprietary Claude Fable 5 and GPT-5.6 Sol models. That is a more useful planning statement than treating one benchmark as a universal win.

Kimi reports 67.3 on DeepSWE with mini-SWE-agent and 90.4 on BrowseComp without context management. Harnesses, tools, token budgets, and retry policies differ across vendor evaluations, so validate K3 on your own repository and agent loop.

Use K3 when a 1M context window, very long output, native visual input, and automatic caching are valuable. Route simpler work to a cheaper model when $15/M output would dominate the task bill.

Compare it with GPT-5.6 Sol, GLM-5.2 and DeepSeek V4, or Kimi K2.7 Code. You can also model your own token mix with the AI Pricing Calculator.

Official sources checked July 19, 2026: Kimi K3 launch, K3 API quickstart, global API pricing, China API pricing, and API limits.