How do I use Kimi K3 in Claude Code?
Set ANTHROPIC_BASE_URL to https://api.stdcmpt.com, a Standard Compute key in ANTHROPIC_AUTH_TOKEN and ANTHROPIC_MODEL to moonshotai/kimi-k3: Claude Code runs on Kimi K3 without a Moonshot account, on a plan from $19 a month that returns HTTP 402 to new requests once its compute budget is used. Point the haiku slot at a cheaper model: K3 lists at $3 input and $15 output per million tokens, more than GLM 5.3, Qwen 3.8 Max or Gemini 3.1 Pro.
# Kimi K3 for the work, GLM 5.3 Flash for background calls
# No /v1 here: Claude Code adds /v1/messages itself.
export ANTHROPIC_BASE_URL=https://api.stdcmpt.com
export ANTHROPIC_AUTH_TOKEN=YOUR_API_KEY
export ANTHROPIC_MODEL=moonshotai/kimi-k3
export ANTHROPIC_DEFAULT_OPUS_MODEL=moonshotai/kimi-k3
export ANTHROPIC_DEFAULT_SONNET_MODEL=moonshotai/kimi-k3
export ANTHROPIC_DEFAULT_HAIKU_MODEL=z-ai/glm-5.3-flash
export CLAUDE_CODE_SUBAGENT_MODEL=moonshotai/kimi-k3
claude
Without the haiku line, Claude Code sends its background calls (titles, summaries) to K3 as well. The Kimi tab follows Moonshot's Claude Code guide.
Set the background slot before anything else
With a gateway key in ANTHROPIC_AUTH_TOKEN, Claude Code sends its background calls (session titles, summaries, classifiers) to the main model unless ANTHROPIC_DEFAULT_HAIKU_MODEL names another. On K3 that is $3 per million input tokens for housekeeping. The config sends those calls to z-ai/glm-5.3-flash at $0.15, and Moonshot's own setup does the same with its kimi-k2.7-code model.
What works with Kimi K3 in Claude Code
| Claude Code feature | On K3 through Standard Compute |
|---|---|
| Tools and MCP servers | Work, translated to function calls |
| Pasted screenshots | Work on moonshotai/kimi-k3. The US-served variant, moonshotai/kimi-k3:us, is text-only |
| Thinking toggle | No effect. Pick K3's reasoning level (minimal to high) on the Models page |
As with every non-Claude model here, Claude Code's WebSearch tool is unavailable and it assumes a 200K window unless you set CLAUDE_CODE_MAX_CONTEXT_TOKENS.
What a day of Kimi K3 costs
At list prices the $20 Starter budget covers about 3 typical K3 days and Standard ($89 a month, $95 of compute) about 15. Keeping K3 for the main session and a Flash model on background calls is what stretches either.
| Kimi K3 at list price | Cost |
|---|---|
| Typical day | $6.16 |
| Busy day | $63 |
| 20 typical days | $123 |
Kimi K3 list prices per million tokens: $3.00 input on a cache miss, $0.30 on a hit, $15.00 output; a 5-minute cache write costs the input rate. Source. A typical day is 8.5M tokens and a busy one 86.7M, 87.2% of input as cache reads: the median and 90th-percentile coding-agent day in our Agent Burn Index, measured on OpenCode, Cline, Kilo and similar agents, September 13 to October 12, 2026. Priced at the list rates above; your own days will differ.
The catch
If you only want K3 and already have a Moonshot account, its own endpoint, https://api.moonshot.ai/anthropic, works with Claude Code and bills per token. K3 all day is expensive on either route; a plan suits K3 as the model for hard steps next to a cheaper daily driver.
Where Standard Compute fits
No Moonshot account, region choice or prepaid balance: one card payment covers K3 and every other catalog model, with spending capped at the plan's budget.
FAQ
Can I use Moonshot's ID kimi-k3[1m] here?
No. Use moonshotai/kimi-k3 or the short form kimi-k3. The [1m] suffix is how Moonshot's guide tells Claude Code to assume a 1M window; on this endpoint, set CLAUDE_CODE_MAX_CONTEXT_TOKENS instead.
Is K3 better than Claude Sonnet 5.5 for Claude Code?
It costs more per day at list price: about $6.16 against $3.36 for Sonnet 5.5 on the same typical day, mostly because of K3's $15 output and $0.30 cache-hit rates. Try both on a real task; the same key serves both.
Does Kimi K2.7 Code work here?
No, it is not in the catalog. K3 is the Kimi model served; GET https://api.stdcmpt.com/v1/models lists every ID.