kimi-k3 — API, Pricing & Context Window | Vivgrid

kimi-k3 is Moonshot AI's frontier model for coding, reasoning, and agentic workflows. It expands the Kimi family to a 1M-token context window with up to 128K output tokens, allowing agents to reason over large repositories and complete substantial engineering tasks in a single run.

Vivgrid exposes kimi-k3 through the unified, OpenAI-compatible Chat Completions API (/chat/completions). You can use the same API key, client libraries, billing, and observability as the rest of the Vivgrid model catalog.

Specifications

Pricing

Pricing in USD per 1M tokens, matching the provider's rates.

What one request costs

The rates above applied to a 20K-token prompt with a 2K-token reply — a typical single agent turn, small enough that every request is billed at the base rate.

Use kimi-k3 in your coding agent

kimi-k3 works with the same Vivgrid key and https://api.vivgrid.com/v1 endpoint as every other model in the catalog. Grab your API key from the Vivgrid Console, then drop one of the configurations below into your agent of choice.

OpenCode

OpenCode has Vivgrid built in, so you can use kimi-k3 without writing any config: open /models, search for vivgrid, paste your API key and pick the model. The OpenCode tutorial walks through it step by step.

Prefer to configure it by hand? Install OpenCode:

npm i -g opencode-ai

Declare kimi-k3 under the Vivgrid provider in your global OpenCode config, ~/.config/opencode/opencode.json. The @ai-sdk/openai-compatible provider talks to Vivgrid over Chat Completions (/v1/chat/completions):

{
  "$schema": "https://opencode.ai/config.json",
  "model": "vivgrid/kimi-k3",
  "provider": {
    "vivgrid": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Vivgrid",
      "options": {
        "baseURL": "https://api.vivgrid.com/v1",
        "apiKey": "{env:VIVGRID_API_KEY}"
      },
      "models": {
        "kimi-k3": {
          "name": "kimi-k3",
          "tool_call": true,
          "attachment": true,
          "modalities": { "input": ["text", "image"], "output": ["text"] },
          "limit": { "context": 1000000, "output": 128000 }
        }
      }
    }
  }
}

Export your key and launch OpenCode:

export VIVGRID_API_KEY="viv-xxxxxxxxxxxxx"
opencode

Use /models to switch between kimi-k3 and other Vivgrid models mid-session. See the full OpenCode tutorial.

Pi

Install Pi:

npm install -g --ignore-scripts @earendil-works/pi-coding-agent

Add kimi-k3 to the Vivgrid provider in ~/.pi/agent/models.json. Pi talks to Vivgrid over Chat Completions (/v1/chat/completions):

{
  "providers": {
    "vivgrid": {
      "baseUrl": "https://api.vivgrid.com/v1",
      "apiKey": "viv-xxxxxxxxxxxxx",
      "api": "openai-completions",
      "models": [
        {
          "id": "kimi-k3",
          "input": ["text", "image"],
          "contextWindow": 1000000,
          "maxTokens": 128000
        }
      ]
    }
  }
}

Make it the model Pi starts with in ~/.pi/agent/settings.json:

{
  "defaultProvider": "vivgrid",
  "defaultModel": "kimi-k3"
}

Then verify and start Pi:

pi --list-models vivgrid
pi

Or run a one-shot prompt without entering the TUI:

pi -p "explain the auth flow in this repo" --model vivgrid/kimi-k3

See the full Pi tutorial.

API surfaces

The Vivgrid Console catalogs kimi-k3 on one surface, all behind the same API key. The endpoint shown is the one the Console previews for that surface.

Quick start

Get an API key from the Vivgrid Console, then call kimi-k3 directly.

curl https://api.vivgrid.com/v1/chat/completions \
  -H "Authorization: Bearer $VIVGRID_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      { "role": "user", "content": "Say hello in English, Chinese and Spanish." }
    ],
    "stream": true
  }'

Choosing within the Kimi family

Prices are USD per 1M tokens at the base rate. Rows marked this page are the model described here.

Ideal use cases

• Long-context coding agents and repository-scale refactoring
• Complex software engineering, debugging, and code review
• Tool-calling agents that need dependable multi-step reasoning
• Teams seeking a high-capability alternative to frontier US models

Related models

• gpt-6-astra — OpenAI's frontier gpt-6 coding model
• minimax-m3 — a lower-cost coding alternative
• kimi-k2.8-preview — the efficient Kimi sibling, same 1M context at a third of the price
• claude-opus-5-5 — Anthropic's latest flagship coding model