How to judge whether an
AI API relay is worth using
As of 2026-11-04, judge a Claude Code third-party API against seven publicly checkable criteria first: protocol compatibility, how the Base URL is written, whether per-model prices and the billing unit are published, how caching is billed, payment methods and crediting time, documentation and a status page, and whether a small trial is possible. Then four things you can do yourself: read the multiplier correctly, check how caching is billed, reconcile usage against the bill, run a small trial. Send the same fixed input at least three times and drop the first — only then is the reading steady-state.
Updated 2026-11-04
One key for Claude, GPT and leading Chinese models, billed per token, up and running in minutes.
- Billed per token — live rates on /models
- Pay by card (Visa / Mastercard), Apple Pay, Google Pay or crypto
- Self-service: top up and activate instantly
The four quickest checks
Cache needs repeats
A single call tells you nothing about cache hit rate; send the same input at least three times for a steady reading.
Cold start is not steady state
The first call is almost always a cold start; averaging it in skews the hit rate in either direction.
Prices and billing unit
Whether per-model prices and the billing unit are published, and whether the usage block in the response is what you are charged on, are the two items easiest to check when choosing.
Try before you scale
Top up the minimum, run the whole flow, confirm the billing basis and the refund or transfer terms, and only then add volume.
Claude Code third-party API: the selection checklist
As of 2026-11-04, the selection checklist has seven publicly checkable items: ① protocol compatibility - Claude models are served only through the Anthropic Messages protocol, so a relay that offers only OpenAI-compatible endpoints does not serve Claude Code; ② whether the Base URL follows the official documentation (on QCode: https://api.qcode.cc/api for Claude and https://api.qcode.cc/openai/v1 for the OpenAI-compatible protocol); ③ whether per-model prices and the billing unit are published; ④ how prompt caching is billed, whether cache writes and reads count apart; ⑤ payment methods and crediting time; ⑥ whether documentation and a status page are public; ⑦ whether a small trial is possible before you scale. Only after that is the price worth reading: The same phrase "0.1x" can mean 0.1 of the vendor list price or 0.1 of some internal accounting unit; "CNY 1 = $1 of quota" sounds like parity but it is a unit of account, not a discount. Only one number is comparable across providers: what you pay divided by the quota you receive, valued at vendor list price. Work that out first and a lot of price lists become less attractive.
Start by computing your cost per $1 of quota
Take any price list and compute: total actually paid ÷ quota received at vendor list price. Amortise one-time fees, processing fees and any expiry on bonus credit. Bonus credit with a deadline or a withdrawal restriction cannot be counted at face value. What you end up with is a single figure — currency per $1 of quota — that can be compared across providers.
How to sequence a trial
Top up a small amount and run the whole path: sign-up, payment, key issuance, first request. What matters here is not speed but whether anything gets stuck and whether support is reachable.
Send one fixed input at least three times, drop the first, and read the usage block on the rest. Add a unique random string per test so your own cache does not contaminate the reading.
Sum the usage across those calls and reconcile it line by line against the bill. If it does not match, ask. If the answer is not coherent, do not scale up.
What you can measure vs what you cannot
You can measure this
① Your real cost per $1 of quota (arithmetic). ② Steady-state cache hit rate (fixed input, repeated). ③ Whether per-model prices, the billing unit, documentation and a status page are public. ④ Whether usage reconciles with the bill. ⑤ Whether there are daily caps, concurrency limits or per-transaction top-up limits.
You cannot measure this
① The provider's compliance and licensing standing. ② The provider's cash position and how long it will last. ③ Whether your request bodies are retained or reused. ④ Real capacity at peak. These can only be judged from terms, reputation and time — which is why "small first, scale later" beats any single test.
Two common shapes
Pay-as-you-go against a balance
You are charged for actual token consumption. Nothing is wasted and the balance carries over; what to check is the unit-price basis, whether the balance expires, and whether it can be transferred. Suits spiky usage.
Monthly plan with a daily cap
A fixed monthly fee buys a daily or monthly allowance. Cost is predictable; the catch is that unused allowance is usually neither refunded nor carried over, so the nominal unit price should be divided by what you will actually consume, not by the cap. Suits steady usage.
Three details that get overlooked
① Per-transaction top-up caps: payment channels impose them, so large amounts often need several transactions — ask before you buy. ② Whether the balance can be transferred or withdrawn: most digital goods are non-refundable, so transferability decides how much you can salvage if things go wrong. ③ Whether currencies are fungible: on some platforms the local-currency and USD balances are separate ledgers and cannot offset each other.
On QCode
We bill per token; per-model rates are listed on /models. The usage block in the response is the basis for deduction, and the bill can be reconciled line by line. Everything above applies to us as well — we suggest you buy a Small Pack first and run it once before deciding. Two things to state plainly: we do not offer refunds (a provisioned digital good occupies account resources immediately) and Pay-as-you-go Packs are non-refundable once purchased, but member-to-member balance transfers are supported; and the CNY and USD balances are separate ledgers that do not mix. The one exception is upgrading: after buying a higher-quota plan, email hi@qcode.cc and the unused time on the old plan is credited back to your QCode balance pro rata (to your balance, not to the original payment method). As of 2026-11-04 you start in three steps: create an account, pay on the top-up page, then point ANTHROPIC_BASE_URL at https://api.qcode.cc/api and fill in the key that starts with cr_. You can pay with Alipay, WeChat Pay, cards (Visa / Mastercard), Apple Pay / Google Pay and cryptocurrency: cards go through the international checkout and the balance is credited as soon as payment completes; for crypto, it is a manual on-chain transfer verified by hand; send a small test amount first. Setup is on /claude-code-relay-api, limits and usage on /claude-code-third-party-api-limits.
FAQ
Is a 0.1x multiplier the same basis as "CNY 1 = $1 of quota"?
No, and they are not directly comparable — the first is a ratio against vendor list price, the second is a unit of account. Convert both to "what you pay ÷ quota at vendor list price" and you are finally using the same ruler.
Which third-party API is best for Claude Code?
No ranking - the recommendation here is a checklist, not a list of providers. Ask item by item: protocol compatibility (whether it can serve Claude Code at all), whether the Base URL follows the official documentation, whether per-model prices and the billing unit are published, how prompt caching is billed, payment methods and crediting time, whether documentation and a status page are public, and whether a small trial is possible; then redo the arithmetic on your own usage. Check that the Base URL in the documentation matches the one you put in your client. If the documentation shows a path, prefix or suffix that is not in your setup, there is a mismatch — not necessarily bad, but worth asking what exactly it changes.
How do I measure cache hit rate properly?
Fix one reasonably long input, send it at least three times, drop the first (almost certainly a cold start) and read the cache-read figures in the usage block of the rest. Give each variant its own random string so separate tests do not hit each other's cache.
The bill does not match usage — am I being overcharged?
Not necessarily. Common legitimate reasons: cache writes and reads priced separately, thinking tokens counted as output, tool calls adding extra turns, and daily totals cut on local time versus UTC. Establish the basis first, then judge whether there is an error.
How do I assess the risk of a provider disappearing?
Look at whether you can start small, whether the refund and balance-transfer terms are written down, and how quickly support answers. Then look at time — how long the same domain has run steadily carries more information than any marketing copy. Do not commit a large amount up front.
What usually goes wrong when Claude Code talks to a third-party API?
The pitfalls sit in how the endpoint is written and how usage is priced. Claude runs only over the Anthropic protocol, so a relay with just OpenAI-compatible endpoints leaves Claude Code unable to connect; if unit prices and the billing unit are not published, multipliers and internal exchange rates cannot be compared. In addition, for a given model and input there is a range of how fast a response can be. If one call is unusually quick, glance at the cache-read line on the bill. Fast can be good news (a cache hit) — and cache reads are priced apart from cache writes. As of 2026-11-04, on QCode Claude runs over the Anthropic protocol with ANTHROPIC_BASE_URL set to https://api.qcode.cc/api and a key starting with cr_.
Sources
The criteria on this page rely on publicly checkable sources: the protocol and usage-field definitions in each vendor's API documentation and the cache pricing rules on their pricing pages. No specific provider is named here and no unverified rumour is repeated; your own measurements and each vendor's official terms govern. Checked 2026-09-02. Re-checked on 2026-11-04 against three pages on docs.qcode.cc: endpoints and API paths, the Claude Code tutorial and payment methods.
Point the method at us
Start with a Small Pack and run it once: reconcile one bill, read one usage block, then decide whether to scale.
Further reading
Claude Code team shared keys guide
One sub-API key per person, a daily quota per key, qot_ usage queries, offboarding, and how Anthropic Team / Enterprise compare.
Claude Max and Team monthly API credits
Live since 2026-10-07: $100 / $200 / up to $500, claimed into a Console organization, not for Claude Code, no rollover.
Claude Code in mainland China
Networking, endpoints and the usual configuration problems.
Crypto payment guide
Nine coins, 56 networks, zero platform fee for AI coding API credits.
This page offers general evaluation methods and is not an assessment or endorsement of any specific provider. Terms, limits and pricing rules are governed by each provider's own pages and may change at any time.