Auto routing
The auto model routes each request to a complexity tier, minimising cost without giving up quality.
The four tiers
| Tier | Typical request | Routes to |
|---|---|---|
| SIMPLE | Greetings, classification, extraction | qwen-flash |
| MEDIUM | General writing, summarisation | qwen-plus |
| COMPLEX | Multi-step technical work, code | deepseek-flash |
| REASONING | Proofs, deep analysis | qwen-max |
When to use auto vs. pinning a model
- Mixed traffic (chat + code + analysis) → use auto; lowest average cost
- Quality-critical work (contract review) → pin qwen-max
- Cheapest possible (classification, extraction) → pin qwen-flash
- Long context plus code or function calling → pin deepseek-flash (1M context, input from $0.003/1M on cache hits)
The routing decision is transparent: the model field in the response tells you what was actually used, and that is what you are billed for.