How to check whether a “cheap Claude API” is actually Claude
There are now dozens of services selling Claude / GPT / Gemini access at 20–70% below list price. Some are legitimate resellers with volume pricing. Some are reselling stolen keys. Some quietly swap in a smaller model and bill you for the big one.
Price tells you nothing about which one you are dealing with. These three checks do.
1. Make the model report its own name
A real provider returns the actual upstream model identifier in the response body, not just the alias you asked for.
curl -s https://your-provider.example/v1/chat/completions \
-H "Authorization: Bearer $KEY" \
-H "Content-Type: application/json" \
-d '{"model":"claude-sonnet-5","messages":[{"role":"user","content":"hi"}],"max_tokens":16}' \
| jq -r '.model'
If the answer is a date-stamped upstream id (or at least a stable, specific identifier), that is a good sign. If it echoes back exactly the string you sent and nothing else, ask them how they map aliases to upstream models — in writing.
2. Verify pricing per model, before you pay
A legit reseller can publish a per-model price table because their cost structure is per-model. A service that only says “up to 70% off” and asks you to DM for prices is not a business, it is a gamble.
Test it the same way you test anything else: send one small request, then read the billing line in your dashboard and check that tokens_billed × published_price matches what was deducted.
That is a five-minute test and it catches most of the shady ones.
3. Check the failure modes
Ask what happens when:
- the upstream returns 5xx mid-stream
- you cancel a request halfway
- a model is temporarily unavailable
The answers that should make you walk away: “we still charge for failed calls”, “we don’t support cancel”, “we’ll route you to a similar model” (without telling you which one).
A provider that releases the hold on upstream 5xx is telling you something about how they bill, and it is usually the single best predictor of whether the invoice at the end of the month will look sane.
What actually drives the bill
Almost everyone comparing providers compares input prices. Input is the cheap half.
In practice, for chat workloads, output tokens dominate: they are typically 4–10× the input price, and long answers are exactly what you are paying for. Cache reads are the other big lever, and they only exist if the provider supports prompt caching.
So the comparison that matters is:
your monthly mix (input / output / cache)
× each provider's price for those three lines
Not the headline “$x per million” on a landing page.
A quick sanity check you can run right now
Two commands, no signup:
-
GET /v1/models— does the catalog list the model you want, with a price and a status? - One 16-token chat request — does the response report a real upstream model name, and does your balance move by roughly
tokens × published price?
If a provider cannot survive those two, nothing else they promise matters.
Disclosure: I work on Tapix AI, which sells Claude / GPT / Gemini access below list price. So run the checks above on us too — that is the point of having them. Our per-model pricing is public and the free credit does not require a card.
Top comments (0)