How do I use Kimi K3 in Codex CLI?
Start Codex with standardcompute launch codex -- --model moonshotai/kimi-k3, or give ~/.codex/config.toml a Responses provider at https://api.stdcmpt.com/v1 with model = "moonshotai/kimi-k3": Codex runs on Kimi K3 without a Moonshot account, on a plan from $19 a month that answers HTTP 402 once its compute budget is used.
# Install Codex and the Standard Compute launcher once
npm install -g @openai/codex
curl -fsSL https://standardcompute.com/cli/install.sh | sh
# Start Codex pinned to moonshotai/kimi-k3
standardcompute launch codex -- --model moonshotai/kimi-k3
The launcher leaves your own Codex config alone. The Kimi tab follows Moonshot's Codex guide.
What works with K3 in Codex
- Shell commands, file edits and screenshots work; K3 reads images.
- Codex's built-in web search is off on this connection.
- The launcher lists every served model under
/model, so a cheaper model is one command away for routine steps. - Your ChatGPT plan is not used; requests bill to the plan's compute budget.
K3 is the expensive part of a Codex day
A typical coding-agent day on K3 costs about $6.16 at Moonshot's list price and a busy one about $63. The $20 Starter budget covers about 3 typical K3 days; Standard, $89 a month, holds $95. Most of a K3 bill is its $15-per-million output and $0.30 cache-hit rate, so long sessions on K3 alone use budgets quickly.
The catch
Moonshot serves K3 over the Responses API itself, and its Codex guide sets base_url https://api.moonshot.ai/v1 with model kimi-k3 and an env_key of KIMI_API_KEY. If you want K3 and nothing else, that is the direct route, billed per token. A Standard Compute key fits when Codex mixes K3 with cheaper models or GPT on one bill.
Where Standard Compute fits
No Moonshot sign-up or prepaid balance: one card payment in US dollars, and /model in Codex switches to a cheaper model when K3 is more than a step needs.
FAQ
Is there a US-served K3 for Codex?
Yes, moonshotai/kimi-k3:us. It is text-only; the base ID also takes images.
Can I keep my ChatGPT login for GPT and use K3 for volume?
Yes. The launcher sets the Standard Compute provider for one session only; running codex on its own keeps your usual sign-in.