Relay365 pricing for Claude Code: one API proxy, one ledger

A note from the Relay365 duty desk about the tab you keep losing

Dear developer, this is the Relay365 duty desk writing at the small hours with one practical question: where did the money go? If your AI spend is scattered across several panels, several keys and several billing pages, you are not managing a budget, you are guessing. Relay365 is an independent third-party relay run like a round-the-clock duty desk, and its whole idea is simple: every 线路 sits on one table, so you can see what you are paying for before the month closes.

The account is prepaid. There is no monthly fee, and every call is settled on the token counts the upstream reports. Nothing is billed for a call that fails upstream with a 502 or 503, and a disconnected stream is billed only for what was actually used. That last detail matters at 3 a.m., when a half-finished stream is exactly the thing you cannot reconstruct from memory.

New accounts also receive $0.25 in signup credit, so you can put a first small request on the board and watch the ledger move before you commit anything larger. The point of Relay365 is not to make spending feel invisible. It is to make spending legible.

Claude Code 2.1.285 or newer is the only road in

Start with the clients, because this is where most confusion starts. Claude models work only from the official Claude Code and Claude Desktop clients, and Claude Code 2.1.285 or newer is the floor. If your Claude traffic is being sent from somewhere else, it will not arrive, and no amount of dashboard tinkering will change that.

Cursor, Cherry Studio, Cline, OpenCode, any SDK or cURL pair only with the GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and Hunyuan lanes. Those lanes are also where the familiar chat-completion workflow lives, so the practical move is to bind each key to the lane it is meant to serve and let the Relay365 dashboard show you the result of every call.

Gemini is a separate protocol entirely. It runs on its native protocol only, through Gemini CLI, using :generateContent for single replies and :streamGenerateContent?alt=sse for streams, and it cannot go through /v1/chat/completions. Gemini 2.5 Flash is priced on that lane at $0.2389 input, $1.91 output, $0.2389 cache write and $0.0236 cache read per 1 million tokens, so the client choice and the route belong in the same conversation.

What a day of tokens actually costs on the Relay365 table

Here is the part that quiets a worried developer: Relay365 prices the four token types separately, so a modest input rate never hides an expensive output rate. On Claude Fable 5 (wire name claude-fable-5) the rates are $3.79 input, $18.94 output, $4.66 cache write and $0.2917 cache read per 1 million tokens.

• Input tokens are billed at their own per-1-million-token rate on every lane.
• Output tokens carry their own separate rate, and it is usually the largest line on the bill.
• Cache write tokens are priced separately, so stored context is never hidden inside the input rate.
• Cache read tokens have their own rate too, and the usage log reports cache recall next to the charge.

Compare Claude Fable 5 with the official list price of $10 input, $50 output and $1 cache read, and the Relay365 output rate is about 38% of official. Claude Opus 5.5 (claude-opus-5-5) sits at $1.61 input, $7.59 output, $1.90 cache write and $0.0736 cache read, about 38% of the official $20 output rate. GPT-5.6 Sol (gpt-5.6-sol) is $0.3736 input, $2.09 output and $0.0292 cache read, with the output rate about 10% of the official $20.

For example, if you send 1,000,000 input tokens and 100,000 output tokens a day on Claude Fable 5, that is $3.79 plus $1.894, or $5.684 a day and $170.52 across 30 days. Add 1,000,000 cache read tokens a day and the same month adds $8.751. Write those three numbers on the desk whiteboard and the bill stops being a surprise.

Cache read is the quiet lever in most production workloads, and Relay365 shows it as its own line rather than folding it into input. On Claude Fable 5 a cache read is $0.2917 per 1 million tokens against $3.79 for fresh input, so a prompt that repeats across calls can cost less on reads than on new tokens. The dashboard's cache recall column lets you check whether that is actually happening for your traffic.

值守卡 duty passes for the Claude and GPT lanes

Some months deserve a 值守卡 duty pass instead of wallet traffic. The Claude pass runs 14 days with $700 of quota, a cap of $380 per week, and a price of $50.97 for 14 days. Quota is drawn at the duty-pass rate x2.5, which is at most about $280 of usage at list prices, and less on models whose pass rate sits above list. The pass comes with a dedicated key that never touches the wallet.

The GPT pass runs 30 days with $3000 of quota, a $900 per week cap and a price of $81.53 for 30 days, with at most about $1200 of usage at list prices under the same duty-pass rate x2.5. Because the pass key is separate, you can tell at a glance which traffic is covered by the card and which is drawing down wallet balance.

One more desk note on money: within 14 days of a top-up, unused balance can be refunded minus what was consumed and the payment channel's fees, while bounty and promotional credit are not refundable. 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.

One table, one ledger, one traceable usage log

Everything lands in the dashboard. API keys carry tags, each key is bound to a lane, and quick-wiring snippets sit next to the settings so a new route takes minutes rather than an afternoon. The usage log holds the latest 200 requests, showing the time, model, input, output, cache recall, charge and result of each call. It does not show which key made the call, so keep your own naming discipline.

A ledger covers the wallet, bounty and pass accounts, in-site messages reach the desk once you log in, and the interface speaks six languages. There are limits worth knowing: HTTP 429 appears with more than 120 calls in 60 seconds on one key, or more than 8 requests in flight for one account. Those are conditions to design around, not surprises to diagnose at dawn.

That is the whole shift, from a developer who loses track of AI spend to a developer who can read it: one table for every 线路, one ledger for every dollar, one usage log for every request, and Relay365 keeping watch at the desk while you sleep. When you are ready, create a Relay365 account and start with a small request.