Kimi K3.1 Pro API: pricing and how to use it

Kimi K3.1 Pro (kimi-k3.1-pro), which Moonshot AI released on October 23, 2026, is now available on FastMetal through the API and in chat. It is Kimi K3 post-trained with substantially more compute, and it costs two-thirds of what K3 does.

What Moonshot claims

• The strongest K3-class model. Moonshot calls it its most capable 3T-class open-weight model and the first of a new K3.1 series. The weights are due by November 4, 2026.
• Coding. 65.7% on Terminal-Bench 4.0. Moonshot highlights game development in Unity, Unreal Engine and Godot, and long-horizon programming work.
• 3D, CAD and finance. 92.7 on BenchCAD with tools (Blender and CAD scripts) and 64.4 on the OfficeQA Pro v2 financial-analysis benchmark.
• Fewer factual errors. On Moonshot's own hallucination benchmark (Kimi Hallu Bench, hard set, with tools), the claim-level error rate fell from Kimi K3's 23.4% to 8.3%.
• Cyber capability held back on purpose. Moonshot removed cybersecurity data from training so the open weights stay behind frontier closed models in that area.

These are Moonshot's numbers, not ours.

What we verified

• Image input works. We checked it on a real image before launch.
• Tool calling works, including a forced tool_choice.
• 1M-token context. Served at 1,048,576 tokens.
• Not supported here: PDF input, and strict json_schema output. The model returned prose instead of the schema on 3 of 3 attempts, so we do not advertise it.
• It reasons before every answer, and reasoning tokens bill as output. Asked only to "say ok", it produced about 80 output tokens.

Pricing

$2.11 per 1M input tokens and $10.55 per 1M output tokens, with cached input at $0.211 per 1M. Kimi K3 is $3.17 and $15.83, so K3.1 Pro is two-thirds of the price on both meters. Billing is prepaid, so you can never spend more than you have loaded. Exact current rates are on the model page and pricing.

The host has not published a training policy for prompts, so the model page lists training use as unknown.

Calling it

FastMetal's API is OpenAI-compatible. An example with a tool:

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3.1-pro",
    "messages": [{"role": "user", "content": "What is the weather in Osaka right now?"}],
    "tools": [{
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get the current weather for a city",
        "parameters": {
          "type": "object",
          "properties": {"city": {"type": "string"}},
          "required": ["city"]
        }
      }
    }]
  }'

Because it thinks first, hard prompts take a while. On the API, stream the response ("stream": true). If you leave out max_tokens, FastMetal allows up to 32,768 output tokens on this model.

To try it without code, pick Kimi K3.1 Pro in chat. You can attach an image there.

Compare real answers first

The model page shows Kimi K3.1 Pro's actual answers to our 43 shared prompts: summarization, bug fixing, translation and more. Check them against Kimi K3 and the models you already use.

Browse the full model catalog.