Five Checks Before Buying an AI API Relay
Compare exact models and route classes, inspect real request samples, check subscription terms, and reconcile balances before prepaying a Claude or GPT relay.
Compare the same model and route class
Input, output, cache and subscription quotas have different billing rules. A pooled or reverse route is not directly comparable to an official direct route merely because the model name matches.
Know what the evidence does not prove
The request samples below come from our paid test accounts. A completed API call does not establish upstream model identity or reconciled wallet charges.
Five checks before buying
1. Choose the exact model and client
A Claude Code success does not prove that a generic SDK works. Record the model ID, protocol and intended client before comparing providers.
2. Compare the same route class
Read the provider's route group and separate input and output rates per million tokens. A low-cost account pool must not be compared as an official direct route.
3. Inspect successful and failed calls
Use model-level sample counts, dates and end-to-end response times. One success proves only that one request completed; samples from different hours are not a controlled uptime study.
4. Check the real payment terms
A full-use subscription equivalent is not your bill. Confirm the eligible route, daily quota, expiry and what happens if you use less than the allowance. Start with a small balance.
5. Reconcile before prepaying
Log balance before and after a fixed request alongside input, output and cache tokens. Without those records, do not label the published rate as an observed charge or claim the model's identity is verified.
What our six paid test accounts actually cover
Successful / total recorded requests for the exact model, including manual and scheduled tests. Missing samples stay blank. These test-key results are not proven to use the same upstream group as a quoted price.
| Provider | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|
| UU API | 1 / 3 | 3 / 3 |
| H API | 3 / 3 | 3 / 3 |
| boxying | 2 / 11 | 3 / 3 |
| LinksAPI | 4 / 4 | 3 / 3 |
| OpenOx | 3 / 3 | 4 / 4 |
| APINebula | 3 / 4 | No sample |
The September report explains the test window, latency definition, failures and price sources. A passing API request does not prove model identity or accurate billing.
Frequently asked questions
Does one successful API call prove the model is genuine?
No. It proves only that the tested key, model ID, protocol, and route returned a response at that time. Model identity requires separate capability and consistency evidence.
Why can a provider website be online while its API fails?
The website and model gateway are different services. We therefore keep website reachability, model requests, pricing, and billing as separate evidence.
Is a subscription price the same as the actual token cost?
Only when the eligible quota is fully used under the stated route multiplier. Daily limits, expiry, unused quota, and cache billing can make the effective cost higher.
Can pooled or reverse routes be ranked against official routes?
They can be shown side by side only when clearly separated. They should not compete in the same price ranking because upstream source, client limits, and reliability differ.
Can one API key access every model sold by a provider?
Not necessarily. Keys may be bound to a product line or route group, and some models work only through Claude Code, Codex, or another designated client.