Claude API Proxy and OpenAI-Compatible Endpoints (2026): Claude Code Only for Now
What a Claude API gateway is, why Claude models on cloudgpu.app currently work through Claude Code only, how the OpenAI-compatible endpoint serves the other models, and how to evaluate a gateway before you trust it.
Updated 2026-10-10. Claude models on cloudgpu.app are currently available through Claude Code only. They cannot be called through the OpenAI-compatible endpoint, the OpenAI SDK, curl, Cursor, Cline or Continue right now. The OpenAI-compatible endpoint keeps serving every other model. The examples below reflect this.
If you have searched for a “Claude API proxy” you have probably found dozens of them, from polished services to anonymous relays that vanish after a month. They all promise roughly the same thing: Claude, cheaper or easier to pay for. This post explains what such a gateway actually does, what you can and cannot do with ours today, and the trade-offs to weigh before you send it real work. We run one, cloudgpu.app, so we use it as the worked example and are specific about its limits.
What a gateway is (and is not)
A gateway sits between your client and the model. On every request it:
- checks your API key and prepaid balance,
- passes the request on to the model,
- streams the response back to you, and
- counts input and output tokens and deducts them from your balance.
That is all. A gateway does not train or host its own copy of Claude, and it does not make the model smarter or faster. What it can change is price, payment method and convenience. Anyone who claims more than that, such as “official partner” status they cannot show or uptime guarantees with no status page, deserves suspicion.
cloudgpu.app is not official, authorized or affiliated with the model vendor. It is an independent gateway that runs in Hong Kong.
What works today
The gateway has two doors, and which one you use depends on the model.
| You want | Use | Base URL |
|---|---|---|
| Claude (Opus, Sonnet, Fable) | Claude Code (CLI or IDE extensions) | https://cloudgpu.app |
| DeepSeek, GLM, Kimi, MiniMax, gpt-oss, FLUX, Whisper and the rest | Anything OpenAI-compatible: the OpenAI SDK, curl, Cursor, Cline, Continue, LangChain, OpenClaw | https://cloudgpu.app/v1 |
Claude models are currently available through Claude Code only. A claude-* model id sent to /v1/chat/completions, from any SDK, curl or editor plugin, will fail.
Claude through Claude Code
export ANTHROPIC_BASE_URL=https://cloudgpu.app
export ANTHROPIC_AUTH_TOKEN=cgw-sk-...
export ANTHROPIC_MODEL=claude-sonnet-5-5 # or claude-opus-5-5, claude-opus-4-7 ...
export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-sonnet-5-5 # Haiku is not offered
claude
The base URL has no /v1; Claude Code adds the path itself. You can also put these variables in the env block of ~/.claude/settings.json. The full walkthrough is in Claude Code with a cheaper API endpoint.
Every other model through the OpenAI-compatible endpoint
Most tooling speaks the OpenAI chat-completions format, and for the non-Claude models that is unchanged: change the base URL and key, pick a model id.
from openai import OpenAI
client = OpenAI(
base_url="https://cloudgpu.app/v1",
api_key="cgw-sk-...",
)
resp = client.chat.completions.create(
model="deepseek-v4-flash",
messages=[{"role": "user", "content": "Review this diff for bugs."}],
)
print(resp.choices[0].message.content)
curl https://cloudgpu.app/v1/chat/completions \
-H "Authorization: Bearer cgw-sk-..." \
-H "Content-Type: application/json" \
-d '{"model": "kimi-k2.7-code",
"messages": [{"role": "user", "content": "Explain this stack trace."}]}'
The model list and per-token prices are on the /api page.
Claude prices through the gateway, 8 October 2026
USD per 1M tokens, ours / official list. Every Claude model is exactly 50% of the official price.
| Model | Input | Output |
|---|---|---|
| claude-opus-5-5 | $2.00 / $4.00 | $10.00 / $20.00 |
| claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6 | $2.50 / $5.00 | $12.50 / $25.00 |
| claude-sonnet-5-5, claude-sonnet-5 | $1.00 / $2.00 | $5.00 / $10.00 |
| claude-sonnet-4-6 | $1.50 / $3.00 | $7.50 / $15.00 |
| claude-fable-5-1, claude-fable-5 | $5.00 / $10.00 | $25.00 / $50.00 |
No Haiku model is available. Prompt-cache reads and writes are billed at half the official cache rates. The /api page shows the live table.
The trade-offs, honestly
What you gain
- Lower per-token price. Here, half the official rate on every Claude model.
- Payment that works where you are. Alipay, Google Pay, Apple Pay, Visa and Mastercard from $5 in one checkout, or USDT on TRC-20, instead of needing a card your bank approves for a US SaaS charge.
- One prepaid balance for Claude in Claude Code and every other model in your own tools. It does not expire, which also makes it a hard spending cap.
What you give up
- Client choice for Claude. Claude models currently work through Claude Code only. If your workflow needs Claude inside Cursor, Cline, Continue or your own SDK code, this gateway does not cover that right now.
- Vendor-account features. No vendor console, workspaces or vendor rate-limit tiers. You manage keys and usage in the gateway’s own API console.
- An extra dependency. Your requests now depend on the gateway. cloudgpu.app has no automatic failover today and offers no SLA.
- Trust. Your prompts pass through a third party. cloudgpu.app does not store prompt or completion bodies and keeps token counts for billing, but you should decide whether that is acceptable for your data, as you would with any processor.
How to evaluate any Claude gateway
Whether you pick us or someone else, check these before topping up seriously:
- Is there a concrete per-model price table, dated, next to the official price? “Up to 90% off” with no table is a red flag.
- Does it say plainly which clients work and whether it stores your prompts?
- Does it claim to be official without evidence? Independent gateways should say they are independent.
- Does the model you asked for answer like that model? Run a few of your own prompts and compare against what you expect.
- Does usage match billing? Send a known request and check the tokens charged in the console.
- Can you start small? A gateway that pushes large prepayments is asking you to carry its risk.
For sign-up and payment, see Cheap Claude API access with Alipay and no US card. For editors, see Claude in Cursor, Cline and Continue, which explains what to use there instead.
Start with a small top-up
Top up a small amount (a first top-up of $10 or more gets a $5 bonus), run Claude Code against https://cloudgpu.app for a day, compare output quality, latency and billed tokens against what you expect, and only then decide whether it deserves more of your work. Get a key at cloudgpu.app/api.
Try cloudgpu.app — no credit card required
No credit card required. Per-minute billing, deploy in 60 seconds.