Six dimensions, scored independently. We publish our method and our own numbers, including where they are unflattering.
Each dimension is measured, not estimated. Refusal handling uses the same 30-prompt battery described in our method page.
| Dimension | Weight | What we measure |
|---|---|---|
| Refusal handling | 30% | Pass rate on the fixed 30-prompt battery |
| Model coverage | 20% | Number of current frontier models, and whether unrestricted variants exist |
| Latency | 15% | Time to first token and sustained throughput, measured over 100 calls |
| Payment & access | 15% | Whether local payment methods work without a foreign card |
| Support | 10% | Response time and resolution on a scripted set of questions |
| Stability | 10% | Uptime and incident communication over 90 days |