Claude Code API relay: when the Relay365 balance hits zero
One working day, told twice: the version with too many tabs
Start the day the way many developers start it: one browser tab per vendor, one key per vendor, one billing story per vendor. You open Claude Fable 5.1 (wire name claude-fable-5-1) for a writing task, GPT-6 Astra (gpt-6-astra) for a code review, and Gemini 3.1 Pro (gemini-3.1-pro) for something else, and every request carries its own paperwork. Rates are per 1 million tokens in approximate USD, the account itself is kept in CNY, and the four token types — input, output, cache write, cache read — are priced separately on every lane.
Relay365 is an independent third-party relay run like a round-the-clock duty desk. The point of the desk is not to add another tab but to retire several: every lane sits on one table, so one wallet, one set of key habits and one usage log cover the whole shift. For example, if you send 200,000 input tokens and 50,000 output tokens of Claude Fable 5.1 a day, the input side is $0.80 and the output side is $1.0005, so about $1.80 a day at the listed rates. The same morning on GPT-6 Astra, 500,000 input tokens and 100,000 output tokens come to $0.3722 plus $0.372, or about $0.74.
Those figures are worth a second look, because Relay365 quotes Claude Fable 5.1 output at $20.01 per 1 million tokens against an official list price of $50 — about 40% of official. GPT-6 Astra output is $3.72 against $50, about 7% of official. The pricing page shows the full table; these two examples simply show how the desk keeps the arithmetic in plain sight instead of behind a calculator tab.
What actually happens when the balance reaches zero
The honest answer starts with the billing model. Relay365 runs on a prepaid wallet with no monthly fee, and every call is settled on the token counts the upstream reports. When the balance reaches zero, there is simply no balance left to draw on, so the practical step is to top up before the next request rather than expect a prepaid account to keep going on credit.
Two details make that moment less surprising. A call that fails upstream with a 502 or a 503 is not billed, so a broken request does not quietly eat the last of your balance. And a disconnected stream is billed for what was actually used, not for the whole planned call, because the token counts upstream reports are what the desk settles.
Top-ups are not a trapdoor either. Within 14 days of a top-up, unused balance can be refunded minus what was consumed and the payment channel's fees. Bounty and promotional credit are not refundable, so treat that kind of credit differently from money you put in deliberately.
Two other sources of funds belong on the same table. The referral bounty pays 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. A new account also starts with $0.25 of signup credit — small, but something to have on hand before the balance reaches zero.
The same day, told from one table
Now tell the same day again, the way it looks when every lane sits on one table. The morning opens in the Relay365 dashboard, where API keys carry tags, each key is bound to a lane, and quick-wiring snippets get you to a first call without hand-assembling headers. The 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, so read it as a shift log rather than a per-key ledger.
Because input, output, cache write and cache read are charged separately, the log shows which of the four token types a request actually consumed. Gemini 3.1 Pro is listed at input $1.54, output $8.94, cache write $1.54 and cache read $0.1542 per 1 million tokens, while DeepSeek V4 Pro is listed at input $1.45, output $4.35 and cache read $0.0486. A cache-heavy request and a long-output request therefore leave different traces in the same table, which is exactly what you want when the balance reaches zero and you need to see where it went.
The dashboard also keeps a ledger of wallet, bounty and pass accounts, in-site messages to the desk once you log in, and six interface languages.
The lanes carry client rules, and they are firm. Claude models work only from the official Claude Code and Claude Desktop clients, with Claude Code 2.1.285 or newer. Cursor, Cherry Studio, Cline, OpenCode, SDK calls 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 cannot be sent through the chat-completions route. Knowing this before you wire things up saves the kind of morning detour that eats a workday.
Duty passes for the Claude and GPT lanes
For Claude and GPT, the desk keeps optional duty passes on a separate shelf, and this is the part of the zero-balance answer that matters most. A pass comes with a dedicated key that never touches the wallet. The Claude · 14 days · $390 · $210 per week pass costs $50.97 for 14 days and draws up to $390 of quota at the duty-pass rate x2.5, which is at most about $156 of usage at list prices, less on models whose pass rate sits above list, with at most $210 of quota per 7 days. The GPT · 30 days · $2350 · $700 per week pass costs $81.53 for 30 days and draws up to $2350 of quota at the duty-pass rate x2.5, at most about $940 of usage at list prices, less on models whose pass rate sits above list, with at most $700 of quota per 7 days.
Because that key never touches the wallet, a pass keeps working at the moment the wallet balance reaches zero. Two rules keep the books straight. 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. And when a pass is nearly used up, each running call holds $3 x the draw rate while an extra parallel call gets HTTP 429 until the running ones finish.
Other lanes always use the wallet, so for them a top-up is the answer rather than a pass.
Keeping the desk informed and the account moving
One more source of HTTP 429 deserves its own line, because on a busy day it can be mistaken for a balance problem: more than 120 calls in 60 seconds on one key, or more than 8 requests in flight for one account. That is a traffic rule, not a billing rule, and the usage log plus the in-site messages to the desk are where you check it.
The duty desk's answer to the original question is short. Relay365 settles every call from a prepaid wallet on the token counts upstream reports, a failed upstream call costs nothing, a disconnected stream costs only what was used, and a top-up can be refunded within 14 days minus consumption and payment channel fees. Keep the wallet funded, read the latest 200 requests when a figure surprises you, and for the Claude and GPT lanes keep a duty pass ready on the shelf. Create a Relay365 account and bring every lane onto one table.