Claude Code API proxy setup and failure billing at Relay365
What a failed upstream call actually costs
A careful buyer asks the uncomfortable question before the sales question: if the upstream provider returns an error, am I still charged? The Relay365 duty desk answers it directly. Relay365 is an independent third-party relay run like a round-the-clock duty desk, and every call is settled on the token counts the upstream reports. When a call fails upstream with a 502 or a 503, it is not billed. You pay nothing for a request that never delivered.
The second question follows immediately: what about a stream that breaks halfway through? A disconnected stream is billed for what was actually used. The tokens the upstream reported are the tokens you pay for, rather than a whole request charged at full size after it was cut off mid-flight. There is no partial penalty, no flat failure fee and no minimum charge stated for an error response.
Relay365 runs on a prepaid wallet with no monthly fee. That matters for the same reason the failure rule does: there is no subscription quietly accruing while a provider is having a difficult night, and there is no separate line item on an invoice for a request that returned a 502 or a 503. The desk keeps the wallet, the bounty and the pass accounts in one ledger, so you can check the charge yourself instead of taking it on faith.
Because every lane sits on one table, the pricing page can be read line by line rather than guessed at. The four token types — input, output, cache write and cache read — are priced separately, so a request that mixes a long prompt, a long answer and a cached prefix is never averaged into one blended rate. Rates are shown per 1 million tokens in approximate USD, while the account itself is kept in CNY.
Reading a working call, token by token
Take Claude Fable 5, whose wire name is claude-fable-5. Per 1 million tokens, input is $3.79, output is $18.94, cache write is $4.66 and cache read is $0.2917. For example, if you send 1 million input tokens and receive 1 million output tokens a day, that is $22.73 a day, or $681.90 over 30 days, before any cache activity. Add 1 million cache writes a day and the figure rises by $4.66 per day; add 1 million cache reads and it rises by $0.2917 per day.
Comparison with list prices is only meaningful where one exists. For Claude Fable 5, the official list price is $10 input, $50 output and $1 cache read, so the Relay365 output rate is about 38% of official. For GPT-6 Astra, whose wire name is gpt-6-astra, Relay365 charges $0.7444 input and $3.72 output per 1 million tokens, with $0.075 for cache read. The official list price there is $10 input, $50 output and $1 cache read, so the Relay365 output rate is about 7% of official. For example, 1 million input plus 1 million output tokens a day on GPT-6 Astra comes to $4.4644.
The cache columns deserve a second look, because they behave very differently from the input column. Cache read sits far below the input rate on both lanes shown here, which matters when the same long context is replayed on every request. Cache write is priced separately again, so you can see what a fresh cached prefix costs before you commit to a caching strategy.
These figures are examples for planning, not a quotation. The pricing page carries the full table for every lane, and the desk would rather you read the row for the model you actually plan to call than rely on a sample calculation. If your workload is unusual, the in-site messages give you a direct line to the desk.
Which client can actually reach each lane
Client rules at Relay365 are firm, and they are part of what you are buying. 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 the GPT, DeepSeek and other non-Claude lanes, so check the client you already run before you wire a key.
Gemini runs on its native protocol only, through Gemini CLI. A single reply uses :generateContent, a stream uses :streamGenerateContent?alt=sse, and that lane cannot go through /v1/chat/completions. This is why the dashboard hands you quick-wiring snippets: a key bound to a lane and tagged with a name, so the endpoint and the key always match.
Rate limits are stated just as plainly. You will see HTTP 429 if you make more than 120 calls in 60 seconds on one key, or if more than 8 requests are in flight for one account. Those thresholds are on the desk list rather than something to discover during a traffic spike.
Duty passes, the wallet, bounty and refunds
Most traffic simply runs on the prepaid wallet. If you want a Claude duty pass, the 30-day option with $1400 of quota and $450 per week costs $94.03, and it comes with a dedicated key that never touches the wallet. Up to $1400 of quota is drawn at the duty-pass rate x2.5, which is at most about $560 of usage at list prices, and less on models whose pass rate sits above list, with a ceiling of $450 per 7 days.
The GPT duty pass works the same way. The 30-day option with $3000 of quota and $900 per week costs $81.53, with up to $3000 of quota drawn at the duty-pass rate x2.5, at most about $1200 of usage at list prices, and $900 per 7 days. Duty passes are optional and cover Claude and GPT only; every other model always uses the wallet, which keeps the choice simple when you run more than one family.
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. Refunds are available 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. A new account also opens with $0.25 of signup credit.
Tracing every request before you top up
The Relay365 dashboard is where the desk keeps its own records. The usage log shows the latest 200 requests with the time, model, input, output, cache recall, charge and result of each call, so a failed call and a billed call sit side by side in the same view. The log does not show which key made the call, which is why the desk suggests one key per lane, tagged so the trail stays readable.
There is also a ledger of wallet, bounty and pass accounts, so balances live in one place instead of three separate dashboards. In-site messages reach the duty desk directly once you log in, and the interface is available in six languages. The API base is a single endpoint, so adding Relay365 to an existing codebase stays a small change rather than a rewrite.
Start with the questions that matter: a 502 or a 503 is not billed, a broken stream is billed only for what was actually used, the four token types are priced apart, and every lane sits on one table. If that is the relay you want, create a Relay365 account and read the full pricing table before your first top-up.