Features
Core
Usage and cost
- Per-request accounting by token, model, and channel, with the billing detail visible in the log
- Input, output, cache-hit, and reasoning tokens counted separately, including cache billing for OpenAI, Azure, DeepSeek, Claude, and Qwen
- Per-call and per-usage billing policies, plus group ratios
- Internal top-ups and quota allocation (Yipay, Stripe)
Authentication and security
- OAuth sign-in: GitHub, Discord, LinuxDO, Telegram
- OIDC single sign-on
- WebAuthn / Passkeys
- Tokens can be scoped by group, model, quota, and expiry
Smart routing
- Weighted random channel selection
- Automatic retry on another healthy channel after a failure
- Per-user model rate limiting
Formats and conversion
Supported request formats and conversion directions are listed in the API reference. The point is that a client speaking one format can reach an upstream speaking another, without per-provider code.
Reasoning effort
Reasoning effort is selected with a model-name suffix, with no change to the request body:
Gemini also accepts an explicit budget, as in gemini-2.5-pro-thinking-128.
Model and task types
Chat, Responses, Realtime, images, audio, video, embeddings, and rerank, plus asynchronous task APIs such as Midjourney Proxy and Suno.