Moonshot AI's Kimi K3.1 Pro (model ID kimi-k3.1-pro) is now live on HopBase, served over the OpenAI-compatible POST /v1/chat/completions endpoint as well as POST /v1/responses and the Anthropic-compatible POST /v1/messages, so the OpenAI SDK, the Anthropic SDK, Claude Code and Codex CLI all connect directly. It is Moonshot's newest flagship in the Kimi K3 series: a 1M-token context window (1,048,576 tokens), up to 131,072 output tokens per response, deep thinking on by default with the trace returned in reasoning_content, text and image input, tool calling and automatic prefix caching. HopBase bills it against Moonshot's official list price — $2 per million input tokens, $0.20 per million cached input tokens and $10 per million output tokens, about a third below Kimi K3. Any key whose groups include it can call it today — GET /v1/models shows what your key can reach.
Updated 2026-11-17: pricing checked against Moonshot's official pricing page; specs and behaviour measured on the HopBase production gateway.
Working in three minutes: create a HopBase console account and a key, add the Kimi K3 Series Self-Hosted group to it, set the base URL to https://api.hop-base.com/v1 and the model to kimi-k3.1-pro; the curl below runs as is.
How to call it
- Three protocols, one key:
POST /v1/chat/completionsandPOST /v1/responseswith base URLhttps://api.hop-base.com/v1;POST /v1/messageswith base URLhttps://api.hop-base.com(no/v1) for the Anthropic SDK and Claude Code. - Use the full model ID:
kimi-k3.1-pro. Read the authoritative list fromGET /v1/models. - Access: the model is in the Kimi K3 Series Self-Hosted group, which also serves
kimi-k3. Ifkimi-k3.1-prois missing fromGET /v1/models, add that group to the key in the console. - Thinking in its own field: the reasoning trace comes back in
reasoning_content(orthinkingblocks on Messages), separate from the answer.
Pricing: at the official list price
| Per million tokens | Kimi K3 | Kimi K3.1 Pro |
|---|---|---|
| Input | $3.00 | $2.00 |
| Cached input read | $0.30 | $0.20 |
| Output | $15.00 | $10.00 |
These are Moonshot's published list prices; there is no tiering by context length. Charges are list price times actual usage, itemised per request in the console, and your effective unit price depends on your group. Thinking tokens are billed as output. Cached reads cost a tenth of the input price, so agent loops and long conversations that re-send the same prefix save the most. For enterprise volume, contact sales from the pricing page.
Specs and what we verified
- Context and output: 1,048,576-token context window, up to 131,072 output tokens per response. A single 40,000-token output completed normally in our tests.
- Long context that actually retrieves: with a single input of about 900,000 tokens, the model returned the key fact buried 60% of the way in, exactly.
- Prefix caching is automatic: re-sending a 25,000-token prefix hit about 99% from the second call on, reported in
usage.prompt_tokens_details.cached_tokens. - Images: send them as base64
data:URLs. Public image links return 400. - Tool calling:
tools,tool_choice(auto/required/none), parallel calls, deep JSON Schemas and multi-turn tool results work in streaming and non-streaming mode, on all three protocols. - Leaner thinking than K3: on the same prompts K3.1 Pro typically spends far fewer reasoning tokens, and responses arrive faster — median time to first token was about 2.3 seconds under sustained load.
Parameter guidance
- The model samples at a fixed temperature.
temperature,top_p,frequency_penaltyandpresence_penaltyare accepted and ignored, so clients with sampling sliders keep working. nmust be 1.- Forcing one specific function with
tool_choicereturns 400 while the model is thinking; use"required"to make it call a tool. max_tokensincludes thinking, so leave headroom for it. For long outputs, stream and addstream_options: {"include_usage": true}to get usage in the last chunk.
Three-minute integration
curl https://api.hop-base.com/v1/chat/completions \
-H "Authorization: Bearer sk-your-key" \
-H "Content-Type: application/json" \
-d '{
"model": "kimi-k3.1-pro",
"max_tokens": 4096,
"messages": [{"role": "user", "content": "Introduce yourself in three sentences"}]
}'
Claude Code connects directly over Anthropic Messages:
export ANTHROPIC_BASE_URL="https://api.hop-base.com"
export ANTHROPIC_AUTH_TOKEN="sk-your-key"
export ANTHROPIC_MODEL="kimi-k3.1-pro"
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=1000000
claude
Codex CLI setup, the full parameter table and FAQ are in the Kimi K3 / K3.1 Pro integration guide.
FAQ
How do I call Kimi K3.1 Pro?
Send an OpenAI-compatible request to POST https://api.hop-base.com/v1/chat/completions with model ID kimi-k3.1-pro. An existing OpenAI SDK only changes the base URL, the API key and the model name. /v1/responses and the Anthropic-compatible /v1/messages work too. Your key needs the Kimi K3 Series Self-Hosted group.
How much does Kimi K3.1 Pro cost?
It is billed at Moonshot's official list price: $2 per million input tokens, $0.20 per million cached input tokens and $10 per million output tokens — about a third below Kimi K3's $3 / $0.30 / $15. Charges are list price times actual usage, and your effective unit price depends on your group.
Can I use Kimi K3.1 Pro in Claude Code or Codex CLI?
Yes, directly. Claude Code uses /v1/messages with base URL https://api.hop-base.com; Codex CLI uses /v1/responses with wire_api = "responses" and base URL https://api.hop-base.com/v1. No local protocol router is needed.
Kimi K3.1 Pro or Kimi K3?
Both have a 1M context window. K3.1 Pro is cheaper on every rate, thinks more concisely and allows longer outputs (131,072 vs 32,768 tokens), so new workloads should start there. Both models are in the same group, so one key calls either and you can compare them on your own prompts.
Why does calling kimi-k3.1-pro return 404?
A 404 model_not_found means none of the groups on your key include the model. Add the Kimi K3 Series Self-Hosted group to the key in the console, and spell the ID in full: kimi-k3.1-pro.
Does it accept images?
Yes, as base64 data:image/...;base64, URLs in image_url. Public https:// image links return 400, so download and encode the image first.
Start calling it: create a HopBase console account and a key; full parameters are in the Kimi K3 / K3.1 Pro docs.