Kimi family

Moonshot AI’s Kimi models, including Kimi K3, K2.6, and the coding-focused K2.7 Code. Called through the OpenAI-compatible API; the Anthropic native API is also supported, so they can be used in Claude Code.

Common tasks

Models

For the full model list and live prices, see the models page . Model IDs have the prefix moonshotai/, which can be omitted.

Kimi models are provided by Moonshot AI directly (moonshot), and some models also have providers such as Azure AI Foundry and Alibaba Cloud Model Studio; the provider is selected automatically by default. To specify one, see Provider routing.

Examples

import os
from openai import OpenAI
 
client = OpenAI(base_url="https://api.ofox.ai/v1", api_key=os.environ["OFOX_API_KEY"])
 
resp = client.chat.completions.create(
    model="moonshotai/kimi-k3",
    messages=[{"role": "user", "content": "Explain idempotency in one sentence."}],
    extra_body={"thinking": {"type": "disabled"}},
)
print(resp.choices[0].message.content)

curl https://api.ofox.ai/v1/chat/completions \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/kimi-k3",
    "messages": [{"role": "user", "content": "Explain idempotency in one sentence."}],
    "thinking": {"type": "disabled"}
  }'

Thinking mode

Measured on the same math problem: Kimi K3 took about 30 seconds by default and about 3 seconds with thinking off; K2.6 used about 11,000 thinking tokens and took about 4.5 minutes by default, and about 40 seconds with thinking off. Thinking tokens are billed at the output rate, and their count is in usage.completion_tokens_details.reasoning_tokens.

Reading thinking content: in streaming responses it is in delta.reasoning_content; in non-streaming responses it is in message.reasoning_details, where you concatenate the text of each item in order.

Web search

Add web_search to tools on the Responses API, and the model searches on its own and attaches cited sources to its answer:

import os
from openai import OpenAI
 
client = OpenAI(base_url="https://api.ofox.ai/v1", api_key=os.environ["OFOX_API_KEY"])
 
resp = client.responses.create(
    model="moonshotai/kimi-k3",
    tools=[{"type": "web_search"}],
    input="What is the latest stable version of Python?",
)
print(resp.output_text)
for item in resp.output:
    if item.type == "message":
        for part in item.content:
            for a in getattr(part, "annotations", None) or []:
                print("Source:", a.title, a.url)

• Supported by Kimi K3 and K2.6 in our tests. output contains reasoning, web_search_call, and message items in that order, and the cited sources are in the annotations of the message content.
• Searches are billed per call; for rates, see the model page .

Use in Claude Code

Kimi models support the Anthropic native API at https://api.ofox.ai/anthropic:

When called through the Anthropic API, thinking content is returned as thinking blocks. For the full Claude Code setup steps, see Claude Code setup.

Image input and structured output

• Kimi K3 and K2.6 support image input using standard image_url content blocks; see Vision for the format.
• response_format supports json_schema and json_object; tool calling uses the standard tools parameter. See Structured output for the format.
• Repeated request prefixes are cached automatically; the hit count is in usage.prompt_tokens_details.cached_tokens.

Common errors

For other general errors, see Error handling.

Official docs

• Kimi Open Platform · Documentation 

FAQ