Relay365 GPT API proxy pricing: cache write, cache read, plain tokens

Cache write and cache read, explained at the duty desk

Here at the Relay365 duty desk, one question arrives more often than any other: when the bill says cache write or cache read, what is actually being counted? The short answer is that Relay365 keeps four token types on separate lines of the invoice — input, output, cache write and cache read — and charges each one at its own rate. Nothing is bundled into a mystery line, and nothing is hidden behind a single blended price.

Input tokens are the prompt material you send: instructions, context, retrieved documents, conversation history. Output tokens are what the model writes back. Cache write is the category for tokens written into the prompt cache so that repeated context can be reused later. Cache read is the category for tokens served back out of that cache on a later request. So a cache write is the setup step, and a cache read is the reuse step.

The honest part of this answer is that these two categories do not always move in the same direction. On Claude Fable 5, for example, per 1 million tokens in approximate USD, input is $3.79, output is $18.94, cache write is $4.66 and cache read is $0.2917. There, a cache write costs more than plain input, while a cache read costs dramatically less. On Gemini 3.1 Pro, the shape is different: input is $1.54, cache write is $1.54 and cache read is $0.1542. Same four categories, different economics per lane.

Every lane on one table, every request traceable

Relay365 is an independent third-party relay, and the desk treats pricing the way a duty desk treats logs: one table, nothing hidden. Every model you can reach through Relay365 is listed with its input, output, cache write and cache read rates, so you can compare a lane before you wire it up rather than discovering the shape of a bill afterwards. The pricing page shows the full table; the point here is only that the four token types are always priced separately.

To make the arithmetic concrete, for example, if you send 1,000,000 input tokens and 1,000,000 output tokens a day on Claude Fable 5, that is $3.79 plus $18.94, or $22.73 a day. Add 1,000,000 cache write tokens on the same day and you are at $27.39. If that same prompt context is served from cache instead, the read is $0.2917, so the day lands at $23.0217. The cache write is the bill you pay to make later reads cheap; the cache read is where the saving shows up.

Traceability matters just as much as the rate. In the Relay365 dashboard, the usage log shows the latest 200 requests with the time, model, input, output, cache recall, charge and result of each call, so you can line up a spend with the request that caused it. One honest caveat: the log does not show which key made the call. Keys can carry tags and be bound to a lane, and quick-wiring snippets are provided, but if you need per-key attribution inside that log, it is not there.

Around the log, Relay365 keeps a ledger for wallet, bounty and pass accounts, plus in-site messages to the desk once you log in. The interface is available in six languages.

The trade-offs worth knowing before you send traffic

First, client compatibility is strict, and pretending otherwise is the fastest way to waste a day. 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, any SDK or cURL pair only with the GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and Hunyuan lanes. Gemini runs on its native protocol only, through Gemini CLI, using :generateContent for single replies and :streamGenerateContent?alt=sse for streams; it cannot go through /v1/chat/completions.

Second, there are limits you should design around. HTTP 429 appears when you send more than 120 calls in 60 seconds on one key, or when more than 8 requests are in flight for one account. That is a throughput ceiling, not a quality judgement, and it is easier to handle with a queue than to discover at midnight.

Third, billing is prepaid and honest. Relay365 uses a prepaid wallet with no monthly fee, and every call is settled on the token counts the upstream reports. A call that fails upstream with a 502 or 503 is not billed. A disconnected stream is billed for what was actually used, which is fair, but it does mean a dropped connection can still leave a small charge on the ledger.

Fourth, refunds come with edges. 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 those as spend rather than as stored cash.

When a duty pass is worth it, and when it is not

Relay365 duty passes are optional and apply to Claude and GPT only; other models always draw on the wallet. A pass comes with a dedicated key that never touches the wallet, which keeps pass usage cleanly separated from your prepaid balance. The trade-off is the rate multiplier: quota is drawn at the duty-pass rate x2.5, so a pass buys convenience and a ceiling rather than a discount.

Take the Claude pass of 14 days: $50.97 for 14 days, up to $700 of quota, with at most $380 of quota per 7 days. At that multiplier the $700 of quota covers at most about $280 of usage at list prices, and less on models whose pass rate sits above list. If your daily volume is steady and predictable, that ceiling is comfortable. If your traffic is spiky and you mostly pay for output, the same money on the wallet may simply go further.

The GPT pass of 30 days works the same way: $81.53 for 30 days, up to $3000 of quota, at most $900 of quota per 7 days, covering at most about $1200 of usage at list prices and less where the pass rate sits above list. Read both figures before you buy: the quota number and the usage number are not the same promise.

Why the desk is still a good deal

Compare only where a comparison exists. For Claude Fable 5, the official list price is input $10, output $50 and cache read $1, and the Relay365 output rate of $18.94 is about 38% of official. For Claude Fable 5.1, output is $20.01 against an official $50, about 40%. For GPT-6 Astra, output is $3.72 against an official $50, about 7%. Where no official price is quoted for a family, Relay365 makes no comparison at all.

That gap is where the desk earns its keep, and it is not built on hidden charges. You get a prepaid wallet with no monthly fee, four token types priced separately, every request visible in the usage log, and lanes that all sit on one table so you can move between them without relearning your billing. New accounts start with $0.25 in signup credit. 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.

So the trade-off is plain: a cache write can cost more than a plain input token on some lanes, a duty pass draws at x2.5, and the usage log will not tell you which key made a call. In exchange, you get output rates well below the official list prices where a list price exists, separate pricing for all four token types, and a duty desk that answers in six languages. That is a workable bargain for a developer who wants to read the numbers instead of guessing at them — so create a Relay365 account, top up the wallet, and put your own traffic on the log to see how it adds up.