One key for every model you can actually reach
GPT, Claude and Gemini — plus Kimi, DeepSeek, Qwen, GLM and MiniMax — behind a single OpenAI-compatible endpoint, at official vendor list pricing with zero markup. No China-side signup for the Chinese models. Chat, image and video on the same key, the same balance, one invoice.
New accounts get a $1 signup credit; invite a friend and you both get $0.10 (offers end December 31, 2026; API usage only, once per person, non-withdrawable; monthly quota applies). Live pricing is on /pricing.
Three walls between you and the model you want
Aggregators mostly solve the first one. The second and third are where projects actually stall.
One key, many vendors
Stop provisioning separate SDK clients, separate dashboards and separate invoices. One endpoint, one balance, one receipt — OpenAI- and Anthropic-compatible.
Chinese models, no China account
Kimi, DeepSeek, Qwen, GLM and MiniMax without a Chinese phone number or a Chinese entity. USD billing — no RMB settlement needed.
Text, image and video — one balance
Chat plus image and video generation (Kling, Seedance, Vidu and more) behind the same key and wallet. Most gateways stop at text.
Two values change. That's the whole migration.
Keep your existing OpenAI SDK code, your streaming, your tool calls and your response shape. Model IDs are bare rather than vendor-prefixed — fetch the catalog once and swap the strings.
from openai import OpenAI
client = OpenAI(
base_url="https://api.hefu.hk/v1",
api_key="sk-...",
)
resp = client.chat.completions.create(
model="deepseek-v4.1-flash", # bare model ID
messages=[{"role": "user", "content": "Say hello in one word."}],
)
print(resp.choices[0].message.content)Get the current catalog
Model IDs move as vendors release. Pull the live list once and diff it against your config before you shift production traffic.
curl https://api.hefu.hk/v1/models \ -H "Authorization: Bearer sk-..." | jq '.data[].id' | head -20
Failure is a latency problem, not a billing one
- Circuit breaker, multi-channel routing and cross-model fallback keep a degraded upstream from taking your app down
- Transparent per-token billing with daily usage statements in the console
- Insufficient balance returns HTTP 402 before the call is dispatched
Official list price. Nothing added on top.
Model prices are anchored to the vendor's published list price. We earn from wholesale spread, not from marking up your bill — and we don't take a fee when you top up.
| Cost component | Typical gateway | HeFu |
|---|---|---|
| Token price | Vendor list price | Vendor list price — zero platform markup |
| Top-up fee (card) | ~5.5% (commonly min $0.80/txn) | 0% — every cent lands in your balance |
| Top-up fee (crypto) | ~5% | 0% |
| Monthly minimum | Varies | None — pay per call |
| Receipts | Provider email receipts | One unified billing dashboard |
Illustrative comparison using the 5.5% card fee published by OpenRouter as of November 2026. Fees and pricing change — verify against each provider's current pricing page before relying on these numbers.
The parts that matter once other people touch the account
Cost control, access control and an audit trail — without a spreadsheet on the side.
Sub-accounts & per-key limits
Give each team or service its own sub-account with its own quota and rate limit. One runaway script can't drain a shared balance.
Usage statements that reconcile
Per-model and per-day usage detail in the console, exportable for finance — one bill across every model family.
Pass-through, never trained on
The relay doesn't store request or response content, and your data is never used for model training. Requests still go to the upstream provider you selected.
Straight answers
Is this an OpenRouter alternative?
It's one of several OpenAI-compatible multi-model gateways, and the right choice depends on your constraint. Where HeFu differs: Chinese frontier models on the same key without China-side registration, image and video generation sharing the same balance, and no platform markup or top-up fee. If you need self-hosting, strict EU data residency, or an observability-first control plane, a self-hosted gateway or a governance platform may serve you better.
Which model IDs do I use?
Bare IDs without a vendor prefix — for example gpt-6-luna, deepseek-v4.1-flash, kimi-k3. Call GET /v1/models with your key for the current catalog; the full list is also on the models page.
Do you charge a markup or a top-up fee?
No. Prices are the vendor's published list price per model with no platform markup, and 100% of a top-up is credited to your balance. Live rates are always on the pricing page.
What payment methods work?
Card payments through Stripe and crypto top-ups (including USDT) are available today. Southeast Asian local wallets (GrabPay, PayNow, FPX) are not yet available.
What happens when a provider goes down?
Requests go through a circuit breaker with multi-channel routing and cross-model fallback, so a degraded upstream doesn't take your app down with it.
Is my data stored or used for training?
The relay is pass-through: request and response content is not stored and is not used for training. Requests are still forwarded to the upstream provider you picked, so their terms apply to that leg of the path.
Is there a free tier?
New accounts receive a $1 promotional API credit (offers end December 31, 2026; API usage only, non-withdrawable, one per person). After that it's pay-as-you-go with no monthly minimum.
Two lines to your next model
Get a key, change base_url and api_key, and keep everything else. Official list pricing, zero markup, zero top-up fees — and the models you can't reach today.
Docs: hefu.hk/docs · Models: hefu.hk/models