RelayVano Worldwide API Relay: Claude Code Limits and GPT Setup

Protocols and gates EN2026-11-03

Ten minutes from signup to a working RelayVano key

RelayVano is an independent relay made for developers worldwide, and the whole site, dashboard and docs come in six languages: 中文, English, Español, हिन्दी, Português, Français. Prices show in yuan on the Chinese pages and in approximate dollars on the others, so you can read the numbers in the language and currency that suits you without translating anything twice.

The setup itself is short. First, create an account; new accounts start with $0.40 of credit, which is enough to fire a test request while you finish the configuration. Next, open the dashboard and copy your API key. Then point your client at https://api.relayvano.com and paste the key into the key field your client already has. That is the whole ceremony: one key, one base address, and no monthly fee anywhere in the picture.

Billing is prepaid balance, pay per token, and the balance never expires. Top-ups go through Stripe checkout, with cards accepted everywhere plus WeChat Pay where your account supports it. Unspent top-ups can be refunded within 9 days, so a small first top-up is a low-risk way to see how your own traffic behaves before you commit to a larger balance.

One key, three routes, and the client rules that matter

One API key reaches every model through https://api.relayvano.com, across three protocol routes: the OpenAI-compatible /v1 route, the Anthropic /v1/messages route, and the native Gemini /v1beta route. You do not juggle separate keys per vendor or per family of models.

The pairing rules are specific, and following them saves an afternoon of confusing errors:

Claude models work only in the official Claude Code and Claude Desktop clients (version 2.1.285 or newer).
Codex, Cursor, Cherry Studio, Cline, OpenCode, OpenAI SDKs and cURL pair only with GPT, DeepSeek and the other non-Claude, non-Gemini models.
Gemini models use the native Gemini protocol only (Gemini CLI); they are not OpenAI-compatible.

In practice: if Claude is your target, drive it from Claude Code or Claude Desktop. If Gemini is your target, use Gemini CLI and the native protocol. If GPT, DeepSeek V4 Pro, Kimi K3, Qwen3.8 Max or MiniMax M3 is your target, the coding clients and SDKs in the second list are the natural fit, because they already speak the OpenAI-compatible shape of the request.

Requests per minute: pace the burst instead of firing it

The first ceiling worth planning around is requests per minute. It is the limit that decides whether a burst of parallel completions is a smooth queue or a wave of rejected calls. RelayVano's current allowance is an operational value that can move, so rather than print a figure here that may be stale by the time you read it, we would rather you see the value that applies to your account: sign in and write to us under Messages in the dashboard, and a logged-in member can ask us directly.

While you confirm that number, three habits keep a burst well behaved:

Spread a batch over a window instead of releasing every request in the same second; a small stagger is often easier to manage than a large queue.
Retry with a backoff rather than immediately, so a rejected request does not return into the same crowded moment.
Keep one key per project or per workload, so the traffic pattern you are tuning is the traffic pattern you can measure.

For example, if you send 200 requests a day, that is an example rather than a measured average for your project. At 200 requests a day, a rate ceiling rarely matters at all; what matters is that the work is spread across working hours and that a retry storm never starts. The same 200 requests concentrated into one minute is a different problem entirely, and that is the scenario the requests-per-minute ceiling governs.

Requests in flight: concurrency and what a disconnect really means

Requests in flight are the calls you have started but not yet finished. With streaming replies, a request can stay in flight for a while, so concurrency and rate are two different dials. A client that opens many long-running streams can reach its in-flight ceiling while still sitting comfortably under its requests-per-minute allowance.

There is one billing behaviour to internalise here: if a caller disconnects mid-stream, the request is still finished and billed on the tokens really used. A dropped connection does not refund the work the model already did. That makes blind retries expensive, because the first attempt may still be running on our side while the second one starts.

A practical approach is to cap your own concurrency, keep a modest number of streams open per worker, and check the usage page before retrying anything that ended abruptly. The usage page lists every request with its charge, so you can see whether the interrupted call was recorded and what it cost before you send it again.

Set a modest concurrency ceiling in your client or worker pool, then raise it gradually while watching the usage page.
Prefer fewer, longer streams over many short ones when your workload allows it.
After an interrupted call, look at the usage page first; that is a straightforward way to avoid paying twice for the same prompt.

Request body size: send what the model actually needs

The third limit is the size of the request body, which is the payload you send: system instructions, chat history, files, tool definitions and any context you paste in. Large bodies cost you twice, once in transport and again in prompt tokens, so trimming the payload is a practical way to reduce spend without changing behaviour.

As with the other two ceilings, we would rather you confirm the current body-size allowance for your account than read a number that may no longer apply. Logged-in members can write to us under Messages in the dashboard and ask for the value that governs the endpoint you plan to use.

Good habits that reduce body size without changing behaviour:

Drop redundant history and keep only the turns that still influence the answer.
Summarise long logs or documents before sending them, rather than attaching the raw material every time.
Keep tool definitions lean, and remove tools the task does not use.
Reuse cached context where it makes sense, since prompt tokens, reply tokens, cache stores and cache hits are priced separately and each shows up on the usage page.

What 200 requests a day looks like on the usage page

To make the economics concrete, here is a worked example rather than a promise. Say you send 200 requests a day, each with 1,000 prompt tokens and a 500-token reply, on Claude Sonnet 5 at $1.27 prompt and $6.36 reply per million tokens. The prompt side works out at $0.254 and the reply side at $0.636, roughly $0.89 for the day, before any cache reads or writes.

The same shape of workload on a lighter model lands differently: Qwen3.7 Plus is listed at $0.1458 prompt and $0.4361 reply, and MiniMax M3 at $0.1472 prompt and $0.5889 reply. On GPT-5.6 Sol, RelayVano lists $0.3736 prompt, $2.09 reply and $0.0306 cache hit against an official list price of $4 / $20 / $0.4 cache hit, an OpenAI promotion stated to run at least until 2026-11-21. Because you pay per token with no monthly fee, choosing the model that fits the task is a practical lever on cost.

Every request shows up with its charge on the usage page, so the numbers above are the kind of line items you can verify rather than estimate. If you bring people in with your link, you keep 16% of every top-up they make, and 4% one level further; earnings can be withdrawn from $5.56 or moved straight into your balance.

Ten minutes is enough to have a key, a working base address, the right client for your model family and a small balance to test with. New accounts begin with $0.40 of credit, the dashboard and docs are available in six languages, and the balance you add never expires. Sign up for RelayVano today, send your first request, and watch the charge appear on your own usage page.