AI API Relay Testing Playbook
Continuously updated. Last update: 2026-08-01
Last year I topped up a relay that everyone in my group recommended. Less than a month later the site was a 502, the group was dissolved, and there was no refund path. Since then I’ve held every provider — including the one I recommend — to two standards: the test method is public, and every recommendation states its limits.
The primary recommendation here is Tokeness — its /api/pricing endpoint publishes input/output/cache prices for every model, which is rare quality in an industry of “official X% off” black boxes.
Articles
- Which Providers Are Worth Using — 5 provider types compared, with a copyable latency test (TTFT/TPS/model consistency)
- Tokeness Deep Review — Chinese models from 5.6x-off official. Legit or not? Worth topping up?
- Relay Recommendations & Reviews — 6 selection points, recommended/neutral/not-recommended tiers, cache-price trap
- Getting Started Guide — DeepSeek/GLM/Kimi/Qwen under one key, registration + API key tutorial
Companion series
Selection decisions: ai-relay-selection-2026 — 5-provider comparison, enterprise procurement, low-price traps, ranking anatomy.