DeepSeek V4 API Without a Chinese Phone Number or US Card: Pay in USDT, Use the OpenAI SDK

DeepSeek V4 is the best value in frontier-class models right now, and the hardest to pay for if you are outside China. DeepSeek’s own platform wants a Chinese phone number for some flows, and its card checkout fails for a lot of international banks. The workarounds people trade on Reddit (a friend’s number, a virtual card, a middleman on Telegram) all break eventually.

cloudgpu.app runs an OpenAI-compatible gateway in front of DeepSeek, GLM, Kimi, MiniMax and a few image and speech models, paid from a prepaid USDT or PayPal balance. This post covers what it costs, how to set it up in your tools, and what the gateway does and does not do.

Prices, 6 September 2026

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

Model	Input (ours / list)	Output (ours / list)	Cached input
deepseek-v4-flash	$0.396 / $0.44	$1.188 / $1.32	$0.198
deepseek-v4-pro	$1.3065 / $1.32	$3.564 / $3.96	$1.3065
glm-5.3-flash	$0.0675 / $0.075	$0.225 / $0.25	$0.0675
glm-5.2	$1.26 / $1.40	$3.96 / $4.40	$0.23
kimi-k3	$2.8642 / $3.00	$14.3212 / $15.00	$2.8642
kimi-k2.7-code	$0.855 / $0.95	$3.60 / $4.00	$0.855
minimax-m3	$0.2814 / $0.30	$1.1055 / $1.20	$0.1407
gpt-oss-120b	$0.135 / $0.15	$0.54 / $0.60	$0.0675

Image generation: flux-1-schnell $0.0015 per image, flux-1-dev $0.0125 per image, both half of the vendor list. Speech and transcription models are on the API page.

Every price sits at or under the vendor’s list. Prices are updated when vendors change theirs; the API page is the source of truth.

Set up in 90 seconds

1. Sign up at cloudgpu.app with email or Google. You get $1.00 of trial credit, enough for roughly 800,000 input tokens on DeepSeek V4 Flash.
2. Open the API console and create a key.
3. Point your client at https://cloudgpu.app/v1.

curl:

curl https://cloudgpu.app/v1/chat/completions \
  -H "Authorization: Bearer $CLOUDGPU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": "Summarise the CAP theorem in two sentences."}]
  }'

Python (openai SDK):

from openai import OpenAI

client = OpenAI(base_url="https://cloudgpu.app/v1", api_key="YOUR_CLOUDGPU_KEY")
r = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Write a haiku about GPUs."}],
    stream=True,
)
for chunk in r:
    print(chunk.choices[0].delta.content or "", end="")

Cursor: Settings → Models → OpenAI API key → paste your key, set the base URL to https://cloudgpu.app/v1, add deepseek-v4-flash as a custom model.

Cline / Continue / OpenClaw: choose the OpenAI-compatible provider, same base URL and key, model id as in the table.

Topping up with USDT

1. Go to the billing page and copy your personal TRC-20 deposit address.
2. Send USDT on the TRC-20 (Tron) network from any exchange or wallet. Minimum 1 USDT.
3. The balance is credited after 20 block confirmations, usually a few minutes. What you send is what lands; we absorb the network cost.

If you would rather not use crypto, PayPal top-ups are available on the same page.

What the gateway does for you

• Failover. DeepSeek’s API has bad days. When the primary upstream times out, the request is retried on the second supplier running the same weights. You see one endpoint.
• One bill. DeepSeek, GLM, Kimi and MiniMax on one balance instead of four vendor accounts, four KYC flows and four payment methods that may or may not work in your country.
• Same balance as GPUs. If you also rent GPUs on cloudgpu.app, everything comes out of one ledger. Prototype on a hosted API, move heavy workloads to a rented RTX 4090 running vLLM, keep one invoice.
• Off-peak pricing on DeepSeek. DeepSeek discounts its API at night in Beijing time; we pass the discount through automatically, so calls between 16:30 and 00:30 UTC are billed at the lower rate.

What it does not do

• It is not free. The trial credit is $1.00, and there is a per-IP free tier on a few utility endpoints for testing without an account. After that you pay per token.
• It does not anonymise you. Sign up needs a real email. Abuse (spam farms, anything illegal) gets a key revoked.
• It is not a custom model host. For that, rent a GPU and run vLLM; the Ollama guide shows the flow.

Which model for what

• deepseek-v4-flash for the bulk of agent and chat traffic. Fast, cheap, 128k context.
• deepseek-v4-pro when Flash is not smart enough: hard reasoning, long multi-step tool use.
• glm-5.3-flash for classification, extraction and other high-volume low-stakes calls. It is the cheapest model on the list by a wide margin.
• kimi-k3 for very long documents (256k context) and when you want a second frontier opinion.
• kimi-k2.7-code for coding agents; it has the code-tuned variant’s tool-call behaviour.
• minimax-m3 for a 1M-token context window at a mid price.

Data and jurisdiction

Requests are forwarded to the model vendor or hosting partner; both are in China or run infrastructure there. Do not send data that your contracts or laws say may not leave your jurisdiction. Our gateway does not train on your prompts and does not store request bodies beyond what is needed for billing and error diagnostics (token counts, model, latency, error code).

Start

Create an API key. Ten dollars of USDT is about 25 million input tokens on DeepSeek V4 Flash, which is more than most side projects use in a month.