Kimi K2 and K3 API Without a Chinese Account: OpenAI-Compatible, Pay by Card or USDT

Kimi is one of the strongest open coding and agent models out of China right now, and if you are outside the country it is awkward to pay for: Moonshot’s own console leans on a Chinese phone number and a card checkout that declines a lot of international banks. This post covers how to call Kimi K2 and K3 through an OpenAI-compatible endpoint on cloudgpu.app, what it costs, and how to wire it into your tools.

Prices, 24 September 2026

Per million tokens. The cloudgpu.app column is what you pay; the Moonshot list column is Moonshot’s own published price for the same model on the same day.

Model	Input (ours / list)	Output (ours / list)	Cached input
kimi-k2.5	$0.855 / $0.95	$3.60 / $4.00	$0.50
kimi-k2.6	$0.855 / $0.95	$3.60 / $4.00	$0.4275
kimi-k2.7-code	$0.855 / $0.95	$3.60 / $4.00	$0.855
kimi-k3	$2.8642 / $3.00	$14.3212 / $15.00	$2.8642

Every row sits under Moonshot’s list. kimi-k2.7-code is the one most people want for agentic coding; kimi-k3 is the frontier model for harder reasoning. The API page is the source of truth and updates when Moonshot changes its prices.

Set up in 90 seconds

1. Sign up at cloudgpu.app with email or Google. There is no signup credit; instead a first top-up of $10 or more gets a $5 bonus, so $10 becomes $15.
2. Top up with USDT (TRC-20) or a card on the billing page. A card works from most countries through Visa, Mastercard, Alipay or a mobile wallet; if your bank declines, USDT always lands.
3. Open the API console and create a key (it starts with cgw-sk-).
4. Point any OpenAI client at the gateway:

curl https://cloudgpu.app/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer cgw-sk-..." \
  -d '{
    "model": "kimi-k2.7-code",
    "messages": [{"role": "user", "content": "Refactor this function."}]
  }'

In Cline, Cursor or the OpenAI SDK, set the base URL to https://cloudgpu.app/v1, the API key to your cgw-sk- key, and the model to kimi-k2.7-code or kimi-k3. Nothing else changes.

What the gateway does and does not do

• Does: speak the OpenAI chat-completions format, stream tokens, and bill per token from a prepaid balance you can also spend on rented GPUs.
• Does not: store your prompts to train anything, require a subscription, or expire your balance. Requests go through our gateway in Hong Kong to DeepInfra in the US (one upstream today, so no failover yet), so this is not a data-residency product — if that matters to you, run the open weights yourself on a rented GPU instead.

One balance for the API and the GPUs

The same top-up pays for both the hosted API and a GPU by the minute. Prototype an agent against kimi-k2.7-code, then rent an RTX 4090 at $0.33/hr or a 5090 at $0.59/hr to run your own model or a Stable Diffusion workload, all from one balance. Start at the API page or the deploy page.