How do I use Kimi K3 in Claude Code?

Point Claude Code at Moonshot's Anthropic-compatible endpoint through the env block of ~/.claude/settings.json. With a Kimi Code membership, set ANTHROPIC_BASE_URL to https://api.kimi.ai/coding/, ANTHROPIC_API_KEY to your Kimi Code key and every model variable (ANTHROPIC_MODEL, the OPUS, SONNET, HAIKU and FABLE defaults and CLAUDE_CODE_SUBAGENT_MODEL) to k3.1-pro or k3.1-pro[1m]. With pay-as-you-go API credits, use https://api.moonshot.ai/anthropic with ANTHROPIC_AUTH_TOKEN instead. Restart Claude Code and check /status.

How much does Kimi K3.1 Pro cost in Claude Code?

On the pay-as-you-go Kimi API, kimi-k3.1-pro costs $2 per million input tokens, $0.20 for cache hits and $10 per million output tokens; the Highspeed variant, about 2 to 3 times faster per Moonshot, costs $4 and $20. A Kimi Code membership replaces per-token billing with a plan quota: Plus/Moderato gets k3.1-pro with a 262,144-token context, Pro/Allegretto and above add k3.1-pro-highspeed and the full 1,048,576-token window. The Go plan has no coding quota.

Why do I get errors after configuring Kimi in Claude Code?

The usual causes, per Moonshot's guide: a model variable left unset, so background tasks or sub-agents request a Claude model name the Kimi endpoint does not know; stale ANTHROPIC_* values in settings.json or ~/.zshrc overriding the new ones; or ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN both set. K3.1 Pro also always thinks, so requests with thinking disabled are rejected.

Are Kimi K3.1 Pro's weights open?

Yes. Moonshot published the weights on Hugging Face as moonshotai/Kimi-K3.1-Pro in the first week of November 2026 under a modified MIT licence: model-as-a-service businesses with more than $20 million in revenue over 12 months need a separate agreement, and products with over 100 million monthly users or $20 million monthly revenue must display 'Kimi K3.1-Pro'. It is a 2.8-trillion-parameter MoE with 104 billion active, shipped in MXFP4.

Quick Answer

How to Use Kimi K3.1 Pro in Claude Code (2026 Setup)

Published: November 9, 2026

The short answer

Claude Code can run Kimi K3.1 Pro by pointing its Anthropic API settings at Moonshot’s Anthropic-compatible endpoint. As of November 9, 2026, there are two routes:

Kimi Code membership (flat plan quota) — base URL https://api.kimi.ai/coding/, model IDs k3.1-pro, k3.1-pro[1m] or k3.1-pro-highspeed.
Kimi API pay-as-you-go (per token) — base URL https://api.moonshot.ai/anthropic, model ID kimi-k3.1-pro at $2 input / $10 output per million tokens.

Set every model variable, not just the main one, or sub-agents and background tasks fail quietly. Moonshot’s own Terminal-Bench 4.0 run of K3.1 Pro (65.7%, against 66.4% for Claude Opus 5.5) used the Claude Code harness, so this is a setup Moonshot tests against.

Which route to pick

Prices from Moonshot’s pricing page and Kimi Code docs on November 9, 2026.

	Kimi Code membership	Kimi API (pay-as-you-go)	Third-party hosts
Base URL	`https://api.kimi.ai/coding/`	`https://api.moonshot.ai/anthropic`	Provider-specific
Auth variable	`ANTHROPIC_API_KEY`	`ANTHROPIC_AUTH_TOKEN`	Provider-specific
K3.1 Pro model ID	`k3.1-pro` / `k3.1-pro[1m]`	`kimi-k3.1-pro`	e.g. OpenRouter `moonshotai/kimi-k3.1-pro`
Price	Plan quota (Go has no coding quota)	$2 in / $0.20 cache hit / $10 out; Highspeed $4 / $20; Batch 60% of standard	Together AI $2 / $0.20 / $10; OpenRouter endpoints from about $0.39 / $7.91 (FP4)
Context	Plus/Moderato 262,144; Pro/Allegretto+ 1,048,576	1,048,576	Varies by host
Best for	Daily coding at a predictable monthly cost	Bursty use, automation, CI	Routing several models through one key

Model prices across vendors are in our current API price table.

Step 1 — Clear old settings

Moonshot’s guide warns that stale values in the env block of ~/.claude/settings.json override variables exported in your shell. Remove any old ANTHROPIC_BASE_URL, ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_MODEL, the ANTHROPIC_DEFAULT_*_MODEL entries, CLAUDE_CODE_SUBAGENT_MODEL, CLAUDE_CODE_EFFORT_LEVEL and CLAUDE_CODE_AUTO_COMPACT_WINDOW, and check ~/.zshrc or ~/.bashrc for leftover ANTHROPIC_* exports.

Step 2 — Add the Kimi variables

For a Kimi Code membership with the 1M window (Pro/Allegretto and above), Kimi’s docs give:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.kimi.ai/coding/",
    "ANTHROPIC_API_KEY": "<your Kimi Code key>",
    "ANTHROPIC_MODEL": "k3.1-pro[1m]",
    "ANTHROPIC_DEFAULT_FABLE_MODEL": "k3.1-pro[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "k3.1-pro[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "k3.1-pro[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "k3.1-pro[1m]",
    "CLAUDE_CODE_SUBAGENT_MODEL": "k3.1-pro[1m]",
    "CLAUDE_CODE_EFFORT_LEVEL": "high",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1048576",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "1048576"
  }
}

On Plus/Moderato use k3.1-pro and 262144 for both context values. For pay-as-you-go, switch the base URL to https://api.moonshot.ai/anthropic, put the key in ANTHROPIC_AUTH_TOKEN instead of ANTHROPIC_API_KEY, and use kimi-k3.1-pro as the model ID (Moonshot’s own example still shows kimi-k3[1m], which selects the older Kimi K3 at $3 / $15). The key sits in this file in plain text, so never commit it.

Step 3 — Verify

Restart Claude Code and run /status: the base URL and model should show the Kimi values. Then send a short message. For the pay-as-you-go route, Moonshot suggests a one-token curl against /anthropic/v1/messages first; a 401 means the key does not match the platform.

Things that behave differently

Always-on thinking. K3.1 Pro only runs with thinking enabled; requests with it disabled are rejected. Reasoning effort accepts low, high and max (API default max).
Fixed sampling. Temperature 1.0 and top-p 0.95 are fixed; omit them.
Missing tiers fail silently. If ANTHROPIC_DEFAULT_HAIKU_MODEL, ..._FABLE_MODEL or CLAUDE_CODE_SUBAGENT_MODEL is unset, Claude Code asks the Kimi endpoint for a Claude model name and those tasks error out.
Support boundary. This is Moonshot’s integration, documented by Moonshot. Claude Code’s interface and variables can change between versions, so re-check after upgrades.

Should you self-host instead?

The weights are open (Hugging Face moonshotai/Kimi-K3.1-Pro, 2.8T MoE / 104B active, MXFP4), but the vLLM recipe targets an 8-way tensor-parallel setup on two GB300 trays, one B300 node or one MI355X node. For one developer, the API or a plan is far cheaper. Hardware detail for the K3 family is in how to run Kimi K3 locally; the model itself is explained in what is Kimi K3.1 Pro.

Last verified: November 9, 2026.