AI API relay reviews: availability and prices
Published rankings
Grok 4.7
| Rank | Relay | Channel group | Availability | Input | Output | Samples | Observation window |
|---|---|---|---|---|---|---|---|
| 1 | Modelflare | grok-stable | 95.83% | CNY 1.4806 | CNY 4.4418 | 48 | 2026-10-17T22:00:00.000Z/2026-10-18T22:00:00.000Z |
| 2 | Cloudbridge | Grok-flex | 93.75% | CNY 1.56 | CNY 4.68 | 48 | 2026-10-17T22:00:00.000Z/2026-10-18T22:00:00.000Z |
| — | CUN.ai | SVIP | No published data | CNY 1.46 | CNY 4.38 | — | not-measured |
| — | Yomi API | Grok 线路 | No published data | CNY 2.148882 | CNY 6.446645 | — | not-measured |
| — | ZNB.ai | znb-grok-heavy | No published data | CNY 0.30 | CNY 0.90 | — | not-measured |
| — | AIHubMix | default | No published data | CNY 14.77 | CNY 44.30 | — | not-measured |
| — | Cubence | base (0-200K tier) | No published data | CNY 13.42 | CNY 40.27 | — | not-measured |
| — | KIE | default | No published data | CNY 5.37 | CNY 16.11 | — | not-measured |
Claude Haiku 5.5
| Rank | Relay | Channel group | Availability | Input | Output | Samples | Observation window |
|---|---|---|---|---|---|---|---|
| 1 | Modelflare | claude-first-topup | 97.92% | No published data | No published data | 48 | 2026-10-17T22:00:00.000Z/2026-10-18T22:00:00.000Z |
How do you choose an AI API relay?
- Choose the model and workload first. Compare the same model, group, currency and billing unit, with input and output prices kept separate.
- Read availability alongside sample counts and observation windows, then open the station and model details for published records.
- Before longer use, check service terms, data handling and refund conditions, and validate your workload using an account you are authorized to test.
API relay review questions
Does a published price mean the relay has been tested?
No. Public quotes and published measurements are separate facts. A price alone does not establish measured availability or a completed evaluation.
Can relay reviews prove which model is running?
Availability and prices do not prove model identity. Neither response style nor a single test is sufficient; read each report’s scope, timing and limitations.