Gemini CLI setup with Relay365 and pay-as-you-go cache reads

Handover note: what the cache recall line on the meter means

Welcome to the shift, colleague. Before you send your first request through Relay365, the duty desk wants to hand you one thing clearly: cache recall is not the same thing as a normal input. When a prompt or a document can be served from a cached prefix, the meter shows those tokens under cache recall, and Relay365 prices them on their own line rather than folding them into input.

That separation matters on a busy day. Every Relay365 lane sits on one table, and the four token types — input, output, cache write and cache read — are priced separately. Cache write is the work of storing the prefix; cache read is the work of recalling it later. Both appear on the usage log next to the charge, so you can see what you were billed for instead of guessing from a single total.

Relay365 is an independent third-party relay run like a round-the-clock duty desk. Rates are quoted per 1 million tokens in approximate USD, while the account itself is kept in CNY, so you can read the meter in dollars without moving your balance around. The pricing page carries the full table; this note is only about the cache recall column and the duty pass options.

Four token types priced separately on one Relay365 table

Take Claude Fable 5, whose wire name is claude-fable-5. Input is $3.79 per 1 million tokens, output is $18.94, cache write is $4.66 and cache read is $0.2917. The cache read rate is a fraction of the input rate, so a request that leans on a cached prefix is charged far less on that portion than fresh input would cost.

For example, if you send 1,000,000 input tokens, 50,000 output tokens and 200,000 cache read tokens a day on Claude Fable 5, the input line is $3.79, the output line is $0.947, and the cache recall line is $0.05834 — about $4.80 for the day, worked out from the Relay365 rates. Add cache writes only when you are storing a new prefix, at $4.66 per 1 million tokens.

Relay365 also lists the official list price beside the Claude and GPT lanes, which makes the comparison easy to check. For Claude Fable 5, the Relay365 output rate is about 38% of official. For GPT-6 Astra, the Relay365 output rate is about 7% of official. Those are output comparisons; the cache read column is quoted separately on the pricing page.

Cache read rates across a few Relay365 lanes

Across the table, cache read stays well below input on every lane that offers it. Claude Opus 5.5 charges $1.61 for input, $7.59 for output and $0.0736 for cache read per 1 million tokens. If your workload recalls 20 million cached tokens a day on Claude Opus 5.5, the cache recall line is about $1.472 a day.

GPT-6 Astra is quoted at $0.7444 input, $3.72 output and $0.075 cache read. Ten million recalled tokens a day therefore come to $0.75 on the cache recall line, and the Relay365 output rate for this lane is about 7% of official. Gemini 3.1 Pro shows $1.54 input, $8.94 output, $1.54 cache write and $0.1542 cache read; five million recalled tokens a day is about $0.771. No official price is quoted for this family, so no comparison is offered.

Two practical notes from the desk. First, cache read only appears when your client actually reuses a cached prefix, so check the usage log after a run rather than assuming it happened. Second, Gemini runs on its native protocol only, through Gemini CLI: single replies use :generateContent, streams use :streamGenerateContent?alt=sse, and the lane cannot go through /v1/chat/completions.

Wallet billing, duty passes and the desk rules

Relay365 uses a prepaid wallet with no monthly fee. Every call is settled on the token counts the upstream reports, so the meter follows real usage. A call that fails upstream with a 502 or 503 is not billed, and a disconnected stream is billed only for what was actually used. A new account also starts with $0.25 in signup credit.

Two duty passes are worth a look if you are running a heavier Claude or GPT workload. The Claude pass covers 14 days with $700 of quota for $50.97, capped at $380 of quota per 7 days, drawn at the duty-pass rate x2.5 — at most about $280 of usage at list prices, less on models whose pass rate sits above list. The GPT pass covers 30 days with $3000 of quota for $81.53, capped at $900 per 7 days, at most about $1200 of usage at list prices.

A duty pass comes with a dedicated key that never touches the wallet, and passes apply to Claude and GPT only; other models always use the wallet. Refunds are available within 14 days of a top-up for unused balance, minus what was consumed and the payment channel's fees. Bounty and promotional credit are not refundable. The referral bounty pays 18% at level 1 and 6% at level 2 of what referred members pay, withdrawable from $11.11 or moved into the wallet.

Tracing calls and the client rules before you switch over

Before you go live, walk the dashboard with me. The usage log keeps the latest 200 requests, showing the time, model, input, output, cache recall, charge and result of each call, so you can trace any single request back to its cache recall line. The log does not show which key made the call, so keep your own naming tidy. API keys carry tags and can be bound to a lane, quick-wiring snippets sit alongside them, and the ledger separates wallet, bounty and pass accounts.

Client rules are firm, so read this line twice. 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 requests pair only with the GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and Hunyuan lanes. Gemini stays on its native protocol through Gemini CLI and is not OpenAI-compatible.

There are two traffic thresholds to keep in mind: more than 120 calls in 60 seconds on one key, or more than 8 requests in flight for one account, will return an HTTP 429. The interface is available in six languages, and you can send in-site messages to the desk once you are logged in if a meter line looks unfamiliar.

That is the handover. Start small, read the cache recall column on the usage log, and let the duty desk know how the meter looks for your workload. When you are ready, create a Relay365 account and take your first lane for a spin.