Is a cheap Claude API legitimate? Six checks before you send money
Price alone tells you nothing about which one you are looking at. These are checks you can run yourself — including against us, and we fail one of them.
Search for cheap Claude API access and you will find endpoints advertising 50 to 80 per cent off Anthropic’s published rates. Some are legitimate resellers with provider credentials. Some are serving you a cheaper model under the expensive model’s name. Some are fronting somebody else’s stolen key and will vanish with your balance. The three look identical from the outside, and the price alone tells you nothing about which one you are looking at.
Here are six checks that do tell you, in the order of how much they cost you to run. The first four take ten minutes and need no trust at all.
1. Ask the endpoint what it serves
GET /v1/models on an OpenAI-compatible endpoint lists what your key can actually reach. Compare that with the landing page.
A gap in either direction is informative. Models advertised but missing means the marketing is ahead of the product. Models present but unadvertised is usually harmless — but it also tells you the catalogue is somebody else’s, resold.
2. Check that the model answering is the model you asked for
This is the substitution check, and it is the one that actually matters, because substitution is invisible in normal use: the response is fluent, the shape is right, and the bill is lower. A study of 28 relays found that 45.83% of endpoints returned a model that did not match the one requested.
Two ways to check, and you want both:
- Read the `model` field in the response. It is free and it catches the careless cases. It is also trivially forgeable, so a correct value proves much less than a wrong one disproves.
- Send a canary prompt. Ask something whose answer differs between models — not “which model are you”, which they answer badly and which a relay can override with a system prompt. The method, with a script takes about five minutes.
Run the canary several times, not once. The interesting failures are intermittent: we have measured endpoints that return the right model most of the time and something cheaper on a minority of calls, which no single request would reveal.
3. Reconcile one bill against the token counts
Send one non-streaming request. Read usage.prompt_tokens and usage.completion_tokens from the response. Multiply by the published per-token rate. Compare with what actually came off your balance.
They should reconcile. If they do not, stop there — everything else you might check is downstream of billing you cannot verify.
4. Measure it before you commit
Availability and latency are cheap to measure and nobody advertises them honestly. Send a request every few minutes for an hour and record what comes back.
Two things to be careful about, both of which we got wrong ourselves before we got them right:
- Do not send them in a burst. Measured on 2026-08-30, four simultaneous requests to the same model failed one in four while the same model probed serially answered 30 of 30. A burst measures the rate limiter and you will publish it as an outage — three ways this goes wrong.
- Decide whether you are counting retries. The official SDKs retry twice by default, so single-attempt availability describes an experience almost nobody has. Measured on the same catalogue on 2026-08-31: 75 per cent of first attempts succeeded, 93 per cent within three. Both numbers, and which to use.
5. Find out who you are extending credit to
Prepaid balance is the part that actually costs money when a relay disappears. You are not buying tokens; you are lending money against future tokens, unsecured, to whoever is behind the site.
So read the terms and look for a named legal entity, a jurisdiction, and a refund position. A relay with no identifiable company behind it is a relay whose failure has no address.
And keep the balance small. The strongest protection is not a promise in the terms — it is topping up in amounts you would not mind losing, which costs you nothing because a one-line base URL change is also how you leave.
6. Ask what happens on the day it ends
Every relay ends eventually — acquired, shut down, or simply abandoned. The question is what that day looks like for you.
| Ask | A good answer sounds like |
|---|---|
| Is my balance refundable? | A stated policy, with a mechanism, not “contact support” |
| How much notice before shutdown? | A number of days written down somewhere |
| How do I leave? | Change one base URL — and it should be true |
| Do you hold my prompts? | A retention period, and preferably a zero |
The last one is worth more attention than it usually gets. Your prompts are your product, your customers’ data, or both, and they are passing through somebody else’s server. A relay that has never thought about retention has not thought about you.
What a legitimate discount actually comes from
A price well under list is not automatically a red flag. There are real mechanisms, and knowing them lets you tell a plausible discount from an impossible one.
- Volume. A reseller aggregating many customers negotiates rates a single customer cannot.
- Regional pricing. Some providers price differently by market, and a relay operating in one can pass that on.
- Batch and cache. Batch pricing halves input and output on some providers; cache hits cost a fraction of fresh input. A relay that routes accordingly has a genuine margin to share.
- Thin margins on purpose. Some operators run close to cost to win the integration, which is a business decision rather than a trick.
What none of those mechanisms produce is a frontier model at a tenth of its cost. If the arithmetic does not work at any plausible wholesale rate, the gap is being made up somewhere — usually by serving you a different model. That is check 2, and it is why check 2 is the one to run first.
Where to go next
- AI API relays, explained — the four kinds of relay, and which one you are looking at.
- Canary-testing an AI API relay — check 2 in full, with a script.
- Prepaid versus postpaid credit — check 5, and what your balance is actually exposed to.
We sell relay access, so treat this list as something to run against us rather than as reassurance from us. Our measured availability is on the status page, our own price audit found our list-price column wrong in both directions, and check 5 is one we currently fail. That is the point: a checklist whose author passes it perfectly is a sales page in disguise.
Figures in this guide were read on the dates shown beside them. Prices change; where a claim depends on a provider’s published price, the link goes to that provider’s own page so you can check it rather than take ours. This guide is reviewed by 2026-11-30.
Check the numbers yourself
Every model on this station, its per-token price and the provider’s published list price are on the pricing page, with no account required to read them.