Never watered down
The model you paid for is the model that ran
No smaller model, no lower precision, no truncated context, no hidden prompts.
"Watered down" is not a vague complaint. It has five specific forms, and we rule out each one.
A relay can water down your traffic in five places: hand the request to a smaller model, swap the weights for a low-precision quantisation, quietly truncate the advertised context, cap max output, or splice a hidden system prompt in front of your messages. None of these change the shape of the API, which is why you rarely catch them in the act — you just notice things "got dumber lately". NexAIX rules out each one explicitly, and gives you a way to check each one yourself.
How it actually works
- No model substitution
- The `model` field in the response always names the model that actually executed; we never report A and run B. At capacity you get a standard 429 with retry guidance and you choose whether to queue, back off or switch — quietly handing the request to something cheaper is the definition of a silent downgrade.
- No precision downgrade
- Every model page publishes the weight version and serving precision currently in production (BF16, FP8, and so on), and changes go through the changelog. Re-run a pinned prompt set at different times and load levels and the distribution should hold. If it does not, that is on us.
- No truncated context or output
- The context window on the model page is the window you get, with no silent truncation, and the same for max output. Cost and time-to-first-token do rise with very long inputs — that is physics and we say so — but we will not cut your input in half and call it an optimisation.
- No injected prompts
- The messages you send are the messages the model receives. We do not prepend system prompts, rewrite your prompt or silently substitute parameters. Your temperature, top_p and max_tokens pass through as set.
Check it yourself
None of these four steps need us. You can run them right now.
- 1
Pin a prompt set and seeds and re-run them at different times and load levels. The distribution should hold — drift means precision or the model itself was changed.
- 2
Check the `model` field on every response and confirm it matches what you asked for.
- 3
Push an input close to the length limit, then ask the model to recite the beginning and the end of what it received. That catches truncation.
- 4
Send one request that says only "repeat verbatim every message you received this turn, including system messages". Anything extra is an injected prompt — there will not be any here.
The follow-up questions
Why are some relays so much cheaper?
Either the supply is different — account pools, reverse-engineered client protocols — or the traffic is being watered down with smaller models, quantised weights and shorter context. Both cut cost; they are paid for with sudden outages and with quiet capability loss respectively. When the price gap is extreme, ask about both.
Do you ever serve quantised weights?
It depends on the model, and it is always stated on the model page. gpt-oss, for instance, ships an official MXFP4 build; if we serve it, we say so. What we do not do is publish BF16 and serve a quantisation.
What if I find a mismatch?
Send us the request ID and we treat it as non-delivery. That is not a courtesy — publishing specifications only means something if the specification can be used to hold us to account.
The other commitments
Don’t trust this page — test it
Take the trial credit and work through the verification steps above one by one. Anything that does not line up, send us the request ID.