The parameters and their weights
| Part | Points |
|---|---|
| Model quality | 22 |
| Free limits | 18 |
| No card, no phone | 13 |
| Commercial use allowed | 9 |
| OpenAI-compatible endpoint | 4 |
| Stability | 4 |
| Editorial rating | 30 |
Seventy points are mathematical, from six inputs in the public dataset. Thirty points are an editorial rating per provider, shown in data/editorial.json with a public note whenever it moves more than 5 points from the default of 15. The internal rubric behind those ratings is not published.
Today the editorial part is neutral. Every provider carries the same 15 of 30 editorial points, so it moves no position: the order is decided by the 70 measurable points. It stays neutral until a public rubric exists.
What the rules guarantee
- Being a partner scores nothing. The score never reads the partner flag. A partner can appear only on the podium (ranks 1 to 3), marked with a visible “Partner” label and a public note; a partner ranked 4 or lower fails the build.
- Missing information is not estimated. A figure the provider does not publish scores zero, or the middle for the two yes/no requirements. Each provider shows how many of the six inputs are confirmed.
- Minimum evidence. Only verified providers inside the freshness window that offer text models, and with at least 4 of the six inputs confirmed, can enter. The list holds up to 10 and is never padded.
- The editorial part cannot outweigh the data. It can move a provider away from 15 only by as much as its confirmed inputs allow.
- Stability is not measured yet. It needs probes over at least 30 days (12 samples); until then it is unmeasured for everyone.
- Deterministic. The same data gives the same order; ties break by the mathematical score, then by name.
Last data update: 2026-10-12. Full text and every constant: docs/methodology.md and scripts/lib/score.mjs.