GPT API proxy setup and prepaid refunds with Relay365

One working day, two shifts, one prepaid wallet

The Relay365 duty desk knows the rhythm of a working day in AI development. It starts with vendors: tabs open, rates compared, one lane for drafts, another for reasoning, a third for the model that finally handles the long context. Then the afternoon arrives and the same developer wants one table, one wallet, one place to look when a bill needs explaining. Relay365 is built for both halves of that day, and the refund window is one of the quieter parts of the design: money you prepaid and did not use is not simply gone.

Relay365 is an independent third-party relay, run like a round-the-clock duty desk. It bills on a prepaid wallet with no monthly fee, and every call is settled on the token counts the upstream reports. That last part matters for refunds, because the desk can subtract exactly what was consumed rather than guessing from an estimate. The result is a wallet you can top up, use, and inspect, with a clear answer to the question that usually annoys people most: what happens to the part I never spent?

The short answer is simple. If you top up and then request a refund within 14 days of that top-up, the unused balance can be refunded, minus what was consumed and minus the payment channel's fees. Bounty and promotional credit are not refundable, so the desk keeps those categories visibly separate. That distinction is worth understanding before you top up, because it tells you which part of your balance is flexible and which part is a benefit you spend rather than recover.

What the 14-day window actually returns

Imagine the morning-after scenario. You added money for a launch week, the launch slipped, and three days later you notice that most of the deposit is still sitting there. Under the Relay365 refund policy, that unused balance can be refunded within 14 days of the top-up. What comes back is not the full deposit: the desk subtracts what you consumed and the payment channel's fees. If you used nothing, the only reduction is the channel's fees. If you used a portion, the used portion is deducted first, and the remainder is what remains eligible.

Bounty and promotional credit are not refundable, so the desk recommends keeping a mental line between topped-up money and credit. A new account receives $0.25 in signup credit. The referral bounty is 18% at level 1 and 6% at level 2 of what referred members pay, and it can be withdrawn from $11.11 or moved into the wallet. That credit can be genuinely useful, but it does not sit inside the 14-day refund window, and treating it as though it does would only create confusion at the worst possible moment.

The practical rhythm is straightforward. Top up when you have a clear plan for the money. Use the wallet as planned. If the plan changes inside 14 days, ask about the unused balance. Outside that window, the top-up is simply part of your working balance. There is no monthly fee to keep justifying the account in the meantime, so a wallet that sits quietly between bursts of work costs nothing to maintain while you decide what to do with it.

Turning vendor juggling into one table of lanes

Now take the same day and shift it into the afternoon, when the juggling ends. The problem at that moment is rarely the number of lanes; it is the visibility. Which key belongs to which lane? What did that last request actually cost? Why was one call more expensive than another? Relay365 puts every lane on one table, and the dashboard is built to answer those questions without a spreadsheet.

The dashboard provides API keys with tags, a key bound to a lane, and quick-wiring snippets so you can move from comparison to a working configuration. A usage log shows the latest 200 requests, with the time, model, input, output, cache recall, charge and result of each call. It does not show which key made the call, but it does show the charge and the outcome side by side, which is what you need when a number surprises you. A ledger covers the wallet, bounty and pass accounts, so the three kinds of money stay legible. In-site messages reach the desk once you log in, and six interface languages are available.

Every request can be traced in the usage log, and the four token types are priced separately: input, output, cache write and cache read. Separate pricing is what makes the log genuinely explanatory. A cache-heavy workflow, a long answer, a large prompt and a fresh write are four different cost events, and Relay365 shows them as four different lines instead of folding them into one opaque figure. When you are deciding whether a lane is worth keeping, you are looking at the same breakdown the desk uses.

Reading the rates with an example, not a promise

Rates are per 1 million tokens in approximate USD, while the account itself is kept in CNY. Because the log separates the four token types, you can compare lanes on the part of the traffic you actually send. For example, if you send 1 million input tokens and 1 million output tokens on Claude Fable 5, that is $3.79 plus $18.94, or $22.73 in usage. On Claude Opus 5.5, the same example is $1.61 plus $7.59, or $9.20. On GPT-6.1 Sol, it is $0.1986 plus $0.9944, or $1.193. These are examples, not quotes for your project, and the pricing page shows the full table.

Where an official list price is published for the family, the comparison is easy to state plainly. The Relay365 output rate is about 38% of the official list price for Claude Fable 5, about 38% for Claude Opus 5.5, about 7% for GPT-6 Astra, and about 10% for GPT-6.1 Sol. The desk prefers to show the arithmetic rather than ask for trust, because a rate card you can check is worth more than a slogan.

Two details protect the bill on the busiest days. A call that fails upstream with HTTP 502 or 503 is not billed. A disconnected stream is billed for what was actually used. Limits also keep the desk orderly: HTTP 429 appears with more than 120 calls in 60 seconds on one key, or more than 8 requests in flight for one account. None of this needs a long explanation. It needs to be written down, traceable, and visible in the log.

When a duty pass replaces the wallet question

Duty passes are optional and available for Claude and GPT only; other models always use the wallet. A pass comes with a dedicated key that never touches the wallet, which changes the refund conversation in a helpful way: pass quota is not prepaid wallet balance, so it is not part of the 14-day top-up refund calculation. You are choosing between two different instruments, and the desk keeps them apart.

The Claude 14-day pass is $50.97 for 14 days. It draws up to $390 of quota at the duty-pass rate x2.5, which is at most about $156 of usage at list prices and less on models whose pass rate sits above list, with at most $210 of quota per 7 days. The GPT 30-day pass is $81.53 for 30 days, drawing up to $2350 of quota at the duty-pass rate x2.5, at most about $940 of usage at list prices and less on models whose pass rate sits above list, with at most $700 of quota per 7 days.

One rule is worth knowing before you rely on a pass under pressure. A call already let through on a pass is drawn in full from the pass even if it runs past the quota or the 7-day ceiling, and the overrun counts against the earliest-expiring card. When a pass is nearly used up, each running call holds $3 x the draw rate, and an extra parallel call gets HTTP 429 until the running ones finish. In other words, the pass behaves predictably under load, and the ledger tells you where the balance stands.

Client rules that keep the setup honest

Relay365 is straightforward about which clients pair with which lanes. Claude models work only from the official Claude Code and Claude Desktop clients, and Claude Code 2.1.285 or newer is required. Cursor, Cherry Studio, Cline, OpenCode, any SDK and cURL pair only with GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and Hunyuan lanes. Gemini runs on its native protocol only, through Gemini CLI, and is not OpenAI-compatible: it uses :generateContent for single replies and :streamGenerateContent?alt=sse for streams, and it cannot go through /v1/chat/completions.

Put those rules next to the refund policy and the shape of the service becomes clear. The lanes sit on one table, every request can be traced in the usage log, the four token types are priced separately, and prepaid money you did not use can come back within 14 days of the top-up, minus consumption and payment channel fees. That is the whole desk philosophy in one afternoon: fewer open questions, more visible numbers, and a wallet that behaves like money rather than a riddle. Create a Relay365 account and bring your lanes onto one table.