Run Kimi K3 and Kimi for Coding in Claude Code

Claude Code already speaks Kimi's protocol

Claude Code sends Anthropic Messages requests to whatever ANTHROPIC_BASE_URL says, and the router at https://router.apitoken.sale answers that protocol for Kimi subscription aliases. No plugin, proxy or fork is involved: you change environment variables, and every session, tier decision and subagent call goes to Kimi instead of Anthropic. Billing moves to your prepaid apiToken.sale balance at a flat 50% below the official Kimi token rates.

The one thing that makes this setup fail silently is Claude Code's internal model map. It keeps separate models for the main session, its Opus/Sonnet/Haiku tiers and its subagents. Setting only ANTHROPIC_MODEL redirects the visible conversation while background paths — title generation, compaction, Task subagents — still carry inherited Claude IDs and break the moment they run.

See also: Run Kimi, Claude, GPT and Gemini in Kimi Code

Pin the endpoint and every model tier

export ANTHROPIC_BASE_URL=https://router.apitoken.sale
export ANTHROPIC_API_KEY=sk-pool-•••
export ANTHROPIC_MODEL=k3
export ANTHROPIC_DEFAULT_OPUS_MODEL=k3
export ANTHROPIC_DEFAULT_SONNET_MODEL=k3
export ANTHROPIC_DEFAULT_HAIKU_MODEL=k3
export CLAUDE_CODE_SUBAGENT_MODEL=k3
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=1048576
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=1048576

claude --model k3

The three ANTHROPIC_DEFAULT_* variables cover Claude Code's tier routing, CLAUDE_CODE_SUBAGENT_MODEL covers Task subagents, and the two context variables raise both the window and the auto-compact ceiling to K3's 1M tokens. Use the bare subscription alias on the Anthropic lane; the scoped GET /v1/models catalog shows the namespaced kimi/* spellings, so check it before pinning an alias into a long-lived environment.

Do not skip the two 1M variables on the k3 alias and do not keep them on a 256K alias. They tell Claude Code how much context it may use before compacting, and a value the served model does not support distorts that decision in both directions.

Match the alias to the session

Figures are per 1M tokens; Kimi caching is automatic, so cache hits and misses bill as separate legs. For a 256K alias such as k3-256k or kimi-for-coding, keep the tier pins exactly as above but omit CLAUDE_CODE_MAX_CONTEXT_TOKENS and CLAUDE_CODE_AUTO_COMPACT_WINDOW. k3[1m] is a compatibility spelling of K3's 1M mode — the router normalizes it to the provider's real k3 wire model, so both forms cost the same.

A practical split: kimi-for-coding as the daily driver for edits and test loops, k3 when the session needs long-context reasoning over a whole repository, and kimi-for-coding-highspeed only when latency justifies exactly double the base token rates.

Full K3 vs Kimi for Coding comparison →

Verify the route, not the model's introduction

1. 01Start a session and run /status. Confirm the Anthropic base URL is apiToken.sale before trusting anything else in the session.
2. 02Send one trivial prompt — "Reply with exactly: connected". A clean answer proves the key, the base URL and the balance in a single round trip.
3. 03Check the scoped catalog before pinning an alias long-term: curl https://router.apitoken.sale/v1/models with your key lists what the key can actually call.
4. 04Exercise a Task subagent once. It is the path most likely to carry an unpinned tier, and you want that failure on the first day, not mid-refactor.

Do not ask the model to identify itself as a verification method. Claude Code's system prompt can make any backend call itself Claude, so the introduction proves nothing about which model is serving the turn — /status and the request path are the evidence.

Reasoning switches are not model selectors

Setting the model slot to none or off disables K3 reasoning; it does not switch you to a different or older Kimi model. Those turns stay on the K3 tariff either way. kimi-k2.6 is not an addressable public model on the router, so typing it selects nothing — use the aliases from the scoped catalog.

K3 supports low, high and max reasoning effort, with high as the default; Kimi for Coding runs with thinking enabled. Reasoning tokens are a subset of output and bill at the output rate — they are never added again as a separate token class, so a thinking-heavy session shows up as output volume, not a surcharge.

What a Kimi session costs on prepaid balance

Every turn is metered per token at the official Kimi rates above, with the flat 50% discount subtracted before the charge touches your prepaid balance. There is no subscription and no seat fee: an idle week costs nothing, and a heavy refactor costs exactly the tokens it consumed at half the official spend. The same balance covers supported Claude, GPT and Gemini models, so a Claude Code session on Kimi draws from the same pool as everything else you run.

• Set a lifetime spending limit on the key and inspect settled usage in the dashboard.
• Default to kimi-for-coding and escalate whole-repository sessions to k3 rather than running everything at K3 rates.
• Reserve kimi-for-coding-highspeed for latency-sensitive loops; its rates are exactly double the base tier.
• Treat a balance-exhausted response as the signal it is — top up and the next request succeeds; retrying changes nothing.

Per-alias Kimi rates and cache legs →

Live catalog of all supported models and prices →

Frequently asked questions

Related guides