OpenAI-Compatible Relays and Reverse Proxies: How to Compare One Before You Trust It (2026)
Aggregator, reverse proxy, official resale — three different things get called a relay, and the differences decide what happens to your key, your prompts and your prepaid balance. Here are the checks that separate them.
1. Three different things share one word
Comparing relays goes wrong at the first step, because people compare model lists — and model lists are the one thing that looks the same everywhere. An aggregator holds accounts with many upstream vendors and routes your request to whichever it chooses, which buys you catalogue breadth and costs you determinism about who served any given call. A reverse proxy forwards to a single upstream, but "a single upstream" may itself be another reseller, and chains of two or three are common and invisible from the outside. Official resale means the operator buys capacity through a licensed cloud vendor's reseller programme and serves it directly, which is the shortest path but constrains the catalogue to what that vendor carries. None of these is inherently dishonest. They simply fail in different ways, and you cannot reason about the failure modes until you know which one you are on.
2. If you only need Chinese open models, stop here
Worth disposing of the easy case first, because a lot of people over-engineer it. If your workload is DeepSeek, Qwen or GLM, use an official-grade Chinese-model platform such as SiliconFlow: first-party inference, CNY billing, mature ecosystem, no intermediary in the path at all. Routing those models through a relay adds a hop and a counterparty and buys you nothing. Relays earn their keep for one specific problem — reaching overseas closed flagships (Claude, GPT, Gemini) that are not in the domestic catalogues and cannot easily be paid for from a local card. Everything below is about that case.
3. Give the incumbents their due
Each of the commonly named services is genuinely good at the thing it was built for, and choosing by slogan rather than fit is how people end up unhappy. Shenma is strong on catalogue breadth — one key against many vendors makes head-to-head evaluation and price-hunting easy. TeamoRouter has put real work into Claude Code and Codex routing specifically, and takes both WeChat and Alipay. API2D has been operating since the early OpenAI-proxy era, which makes migration nearly free for long-time OpenAI users and means a track record you can look up. If your need lands squarely on one of those strengths, that is the right answer and this page is not trying to talk you out of it. What follows is how to evaluate any of them, including us.
4. How many hops? A test you can run in a minute
This is the check almost nobody runs, and it is the most informative one available. Deliberately trigger an upstream error and read what comes back. Send a request that exceeds the model's context window, or name a model that does not exist, and compare the error body against what the upstream vendor's own API returns for the same mistake. A short path passes the upstream error through largely intact — the same error type, the same field names, often the same request id. A long path rewrites it, because each intermediate layer parses the response and re-serialises its own version: you get a generic message, invented error codes, or a 200 with the error stuffed into the message body. The same logic applies to streaming: ask for a stream and watch whether tokens arrive incrementally or in suspiciously large bursts, which suggests a layer buffering the whole response before relaying it. Neither test needs special tooling — curl and a stopwatch are enough — and both are far harder to fake than a marketing claim about being "direct".
5. What happens to your key, and to your prompts
Security questions about reverse proxies usually get answered with reassurance rather than mechanism, so here is the mechanism. You are not handing over your vendor API key — you do not have one; that is the point. The operator holds the upstream account and issues you a key of its own. The consequences follow directly: your prompts and completions pass through the operator's infrastructure in plaintext, whatever transport encryption is in place, because the request has to be re-signed with the upstream credential. So the question is not "is it encrypted" but "what is retained, for how long, and who can read it" — and the answer should be in a published policy rather than a support-chat reply. Three practical consequences: do not send regulated or personally identifying data through any intermediary you have not diligenced; assume prompts may be logged for abuse detection, because they usually are; and rotate the key the operator issued you on the same schedule you would rotate any other credential. If an operator will not say what it retains, treat the silence as the answer.
6. A prepaid balance makes you an unsecured creditor
Nearly every relay is prepaid, and that arrangement deserves to be named for what it is: money you have handed over for a service not yet rendered, with no security and no deposit protection. If the operator shuts down, is cut off by its own upstream, or simply stops answering, your remaining balance is an unsecured claim against an entity that may not be identifiable. This is the single strongest argument for checking whether there is a traceable company entity with complete terms behind the service, and for the unglamorous discipline of topping up in small amounts. A rate a few percent better is not worth concentrating a large balance behind an anonymous operator. It is also why unusually cheap pricing warrants an extra round of scrutiny: flagship inference has a hard floor of upstream cost, and pricing well below it has to come from somewhere — silent substitution, credits of unclear origin, or collecting prepaid balances at a loss.
7. Verify the model before you send volume
The structural temptation for any intermediary is to bill for a flagship and serve something cheaper, and it is invisible on easy prompts. Do not settle it with promises. Keep three to five prompts that actually discriminate between tiers — one long multi-step reasoning problem, one long code generation-and-debug task, one multi-step maths problem — and run them against both a known-good endpoint and the candidate. Push a genuinely long document to see whether the advertised context window is real. Then send 20–50 requests and record time-to-first-token and the failure rate. If assembling that harness is more work than you want, cocodot's downgrade checker is open source with a hosted version at probe.cocodot.co: paste any OpenAI-compatible base URL and key and it runs the probes for you. It is deliberately usable against us as readily as against anyone else — and an operator's willingness to be tested is better evidence than any pledge.
8. Disclosure, and where we are not the answer
This page is published by cocodot, and we are one of the options in the table, so discount it accordingly. Our specifics, stated plainly so you can check them: we are an overseas-registered company serving overseas flagships through licensed cloud vendors' official resale rather than reverse-engineered access, with no second forwarding layer; one key covers Claude, GPT and Gemini; an Anthropic-compatible endpoint is served on the same host so Claude Code connects by setting ANTHROPIC_BASE_URL (beta); top-up is Alipay in CNY; per-token rates for every model are published at /pricing. The honest boundaries: if your models are Chinese open models, an official platform beats us; if you need the widest possible catalogue for evaluation work, an aggregator beats us; and if you require a direct contractual relationship and a vendor SLA, no intermediary is the right answer and you should get a card the vendor accepts. Where we are worth a look is the narrower case of overseas flagships, local payment, and wanting to verify rather than trust. Start with a small top-up anywhere — including here.
Names you will encounter in the Chinese market, and which lane each sits in (verify live details on each vendor's own site)
| Platform | Primary use | Top-up | Overseas models | Chinese models | Path |
|---|---|---|---|---|---|
| Shenma relay API | Broad multi-vendor aggregation | Metered | Yes, broad | Yes | Aggregated relay |
| TeamoRouter | Claude Code / Codex routing | WeChat / Alipay | Yes | Partial | Aggregated relay |
| SiliconFlow | Official-grade Chinese-model inference | CNY | — (not its focus) | Yes, official-grade | Direct (Chinese models) |
| API2D | Long-established OpenAI proxy | CNY credits | Mainly OpenAI | Partial | Relay |
| cocodot (ours) | Overseas flagships + Alipay card issuance | Alipay | Yes, official resale | Yes | Licensed-cloud official resale |