Routing logic
| Task type | Model | Rationale |
|---|---|---|
| Research, competitive analysis | Kimi K2 | Deep reasoning, multi-source synthesis |
| Cold email, content creation | Kimi K2 | Voice matching, quality gate compliance |
| Objection handling, strategy | Kimi K2 | Nuanced reasoning, context awareness |
| Lead scoring | Fast model | Pattern matching, high throughput |
| Quick lookups, summaries | Fast model | Speed matters, complexity doesn’t |
| Routing decisions | Fast model | Lightweight classification |
| Memory compression | Fast model | Summarization at scale |
How routing is decided
The decision happens before the API call based on:- Message complexity — length, number of entities, implied reasoning depth
- Task type — skill runs always use the model specified in their
SKILL.mdconfig - User plan — free plan routes more aggressively to fast models
- Token estimate — very short tasks default to fast even if complex-looking
Credit impact
Kimi K2 costs more per token than the fast model. The routing system keeps your credits efficient by:- Routing memory compression, scoring, and routing decisions to the fast model
- Reserving Kimi K2 for tasks where output quality directly affects business outcomes
- Giving you visibility into model usage per session in Settings → Usage
Each skill’s
SKILL.md file specifies which model it uses. Research and sales skills use Kimi K2. Ops skills like morning briefing and pipeline review use the fast model.