How to tell whether a cheap LLM API is really serving the model you asked for

Cheap API resellers are everywhere, and most are honest. Some are not: they take your request for an expensive model and quietly answer it with a cheaper one. You cannot see this from the price. You can see it in about five minutes with the three checks below.

1. Read the model field in the response

Every OpenAI-compatible response echoes a model field. It should match what you requested, or be the exact upstream version string of that model. It should not be a different model family, and it should not be a cheaper variant.

curl https://tapixai.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ***" \
  -d '{"model":"claude-sonnet-5","max_tokens":32,"messages":[{"role":"user","content":"Say OK"}]}'

Then look at model in the JSON that comes back. Real examples that pass: claude-sonnet-5, claude-haiku-4-5-20251001 (the dated upstream string). A response that says gpt-4o-mini when you asked for claude-sonnet-5 is a substitution, and you should walk away.

Partial flags are a warning, not proof. Some providers map your request to a variant — for instance a "max" name resolving to a "thinking" build, or a specialised variant resolving to the base model. That is not necessarily fraud, but it must be disclosed. We publish ours rather than hiding them. If a supplier will not tell you what your request maps to, treat that as the answer.

2. Do the price arithmetic yourself

Three numbers decide whether a price is plausible: the vendor's own list price, what you were charged, and the token counts in the usage block. Multiply and compare.

list price of claude-sonnet-5:  $2.00 in / $10.00 out per 1M tokens
your billed amount for the call: read usage.prompt_tokens and usage.completion_tokens
expected = prompt_tokens/1e6 * 2.00 + completion_tokens/1e6 * 10.00

If the amount charged is far below that arithmetic, one of two things is true: the provider has genuinely cheaper capacity (possible, and how resellers work), or the answer did not come from that model. The arithmetic alone does not tell you which — check 1 and 3 do.

A useful sanity floor: nobody sells a frontier model at 5% of list. Real resellers buy in volume and pass on part of the saving. When a price is an order of magnitude under everyone else, the product is usually not the product.

3. Compare behaviour, not marketing

Different models behave differently. Ask two providers the same open question with temperature=0 and compare the outputs. Identical word-for-word output from two supposedly different models is a strong signal that one backend is answering for both.

same prompt to provider A and provider B:
"In one short sentence, explain what a token is in a large language model."

Also check the boring things:

Run these checks on us too

That is the point of publishing this: the checks apply to Tapix AI exactly as they apply to anyone else. Our model list names the upstream version strings we actually serve, our price page shows the official list price next to ours for every model, and every response returns the real model value.

We also run this audit against ourselves, and when we find something that does not match, we publish it rather than quietly fixing it: see the changelog.

If any of the three checks fails on our API, email [email protected] with the request ID and we will trace it end to end and tell you what happened — including if the answer is "our upstream substituted it and we did not know".

Why we wrote this

Because "trust us" is not a verification method, and the cheap end of this market has earned its reputation. A provider that will not let you check is telling you something. A provider that publishes its own failing audit is telling you something better.

See the model catalogue · Compare prices to official list · Start with $1 free credit, no card

Last updated: 2026-11-20最后更新:2026-11-20