Methodology
What we check, how we grade, and — just as important — what we refuse to claim.
What we publish
- Verbatim quotes from each provider's published terms, refund and policy pages, always with a capture date and source URL, archived on the day of capture.
- Public registration data (whois): registration date, registrar, privacy-proxy status.
- Our own test results, produced by the same checks that power the verifier, always dated.
What we don't publish
We do not repeat third-party accusations we cannot reproduce ourselves. If a claim about a provider isn't a quote from its own pages, public registration data, or the output of our own dated test, it doesn't appear on this site. Our build pipeline mechanically rejects accusatory language.
Refund grades
| Grade | Meaning (from published terms only) |
|---|---|
| A | Unused balance clearly refundable, stated process and timeline. |
| B | Refundable with minor conditions. |
| C | Partial or discretionary: some paths refundable, others excluded. |
| D | Narrow refund paths plus clauses that can devalue prepaid balances. |
| F | Explicit no-refund clause covering unused funds. |
| N/A | No terms published at all. |
Grades reflect the text of the terms, not our opinion of the operator, and carry the capture date.
Verification checks
The verifier runs four checks against any OpenAI-compatible endpoint, from your browser:
- Reachability & auth — distinguishes an invalid key from an unreachable or browser-blocked endpoint.
- Model identity — asks the model who made it and compares against the vendor implied by the model name.
- Response fingerprint — checks for structural fields the genuine provider returns.
- Latency — time to first token and total time (informational; never affects the verdict).
Verdicts are deliberately conservative: Likely genuine, Suspicious (one or more checks contradict the claimed model — worth investigating, not proof of wrongdoing), or Unreachable / not testable. A browser-based test cannot prove intent; it can only report what the endpoint returned on a given date.
Site reachability monitor
Separately from the verifier above, an automated job on our side checks once a day (one attempt, UTC calendar date) whether each provider's homepage responds over HTTPS. It is website uptime only — it never sends any API request and never touches a key. A result of "Reachable" means our monitor got an HTTP response with a non-server-error status; "Unreachable from our monitor" means it didn't, which can be caused by regional network conditions, temporary outages, or blocking between our monitor and the site, and says nothing about whether the underlying API works. Each review page shows the latest result and the share of daily checks the site has answered since we started watching it. This monitor does not feed the "Verified by us" column or the verifier's checks in any way — those remain a separate, key-holding test that we run and grade independently.
Independence
We sell nothing, run no affiliate links to any listed provider, and accept no payment for grades or placement. If that ever changes for any page element, it will be labeled explicitly on that element.