Cost-aware routing
Cost-aware routing picks which model tier to use for each turn based on task difficulty. Routine turns go to a fast tier; complex reasoning turns route to a frontier tier. Saves cost and latency on routine turns without sacrificing quality on hard ones.
Why it matters
Frontier models are 5-30x more expensive per token than fast/cheap-tier alternatives. For an agent product handling thousands of user turns per day, the cost difference between "always frontier" and "frontier when needed" is meaningful at scale.
The pattern works because many turns are routine. A user saying "hi" or asking "what time is my next meeting" does not need 200K-token context-window reasoning. A small fast model can handle it in 200ms for a fraction of the cost. Reserving the frontier tier for the genuinely hard turns (complex multi-step planning, long-document reading, subtle reasoning) keeps overall economics sane while preserving quality where it matters.
The routing decision can be:
- Heuristic. Length of the user's message, presence of complex keywords, tool-call complexity.
- Model-driven. A small classifier model decides which tier to use.
- User-controlled. Power users opt into "always frontier" for their account.
The honest trade-offs:
- Misrouting. Routing a hard turn to a fast tier produces a worse answer. Catching this requires fallback logic ("if the fast model says it can't do this, retry on frontier").
- Latency variance. Different tiers have different response times. Users sometimes notice the difference.
How Ajeris handles this
Ajeris uses heuristic routing with model fallback. Most user turns route to a fast tier (sub-second response). Heavy turns (planning across multiple tools, long-document analysis, complex reasoning) route directly to the frontier tier. If a fast-tier turn produces an obviously-poor result, the agent retries on frontier.
Users can override the routing in their Soul preferences. Default behavior is invisible to the user; the agent picks.
