Chinese open-weight models now sit at the frontier. As of June 2026 they account for more than 45% of all traffic on OpenRouter, and DeepSeek V4-Pro leads LiveCodeBench (93.5) and Codeforces (3206), ahead of several closed frontier APIs. The models are no longer the question. The question is how you reach them reliably, and pay for them, from outside China.
This guide covers what actually matters when picking an API relay.
Why you need a relay at all
Try to use Chinese providers directly from abroad and you hit account and payment friction: registration often expects a mainland phone number, and billing is built around Alipay and WeChat Pay. A relay holds the upstream accounts for you and exposes a clean, OpenAI-compatible endpoint you can pay for with an international card.
The five things that actually matter
1. Which model *versions* are live, and how fresh
A relay is only as good as its catalog. Check for the current generation, not last year's:
- GLM-5.2 (Z.ai, June 2026): the current #1 open-weight model, with 51 on the Artificial Analysis index, 62.1 on SWE-bench Pro, and a 1M context. If a relay doesn't carry it, that tells you how fast they update. (Deep dive.)
- DeepSeek V4-Pro / V4-Flash (released April 2026): V4-Pro is the reasoning and coding flagship (1.6T total parameters, 49B active); V4-Flash is the fast, ultra-cheap tier (284B / 13B active).
- Kimi K2.6 (Moonshot): a strong agentic and long-context model from the April 2026 wave.
- MiniMax: the M2 series, for cost-efficient general use.
If a relay still tops out at DeepSeek V3 or GLM-5.1, it's behind.
2. OpenAI compatibility (so you don't rewrite anything)
The endpoint should accept the standard OpenAI request shape (/v1/chat/completions, /v1/messages, /v1/models), so your existing SDK works by changing only the base URL and model name. If you have to learn a bespoke API, that's a tax on every integration.
3. How you pay (and whether billing is transparent)
You want an international card (no Chinese phone or Alipay), pay-as-you-go billing (top up a balance, no monthly lock-in), and a per-call transaction log so you can reconcile spend. Watch for relays that bill on opaque "credits" with no breakdown.
4. Routing, latency and uptime
The models are hosted in mainland China. A good relay runs a Hong Kong or edge hub with streaming passthrough so latency stays low for overseas calls, plus multi-channel failover so a single upstream wobble doesn't take you down.
5. Honest pricing vs the official rate
The fairest model is simple: you pay the provider's real metered price (converted to your currency), not a marked-up "credit" rate. On Turiloop, for example, you pay per token close to the providers' own metered rates, currently discounted across every model.
A quick decision checklist
- Does the catalog list GLM-5.2, V4-Pro/Flash, Kimi K2.6? (freshness)
- Is it OpenAI-compatible? (zero rewrite)
- Can you pay with a card, pay-as-you-go, with a transaction log? (no friction)
- Is there a Hong Kong / edge hub and failover? (latency and reliability)
- Is pricing metered, not opaque credits? (no markup games)
If a relay ticks all five, you're in good shape. Turiloop was built around exactly these five: one OpenAI-compatible key for DeepSeek, GLM, Kimi, MiniMax and Claude, card payment, metered billing.