Looking at the latest round of inference benchmarks and routing tools—from Nace AI's Drex Router reporting sub-100ms routing at roughly 1/11 the cost of Opus-only runs ($1.26 vs. $13.44 on a small game build) to the split between $2/$10 tiers like GPT-6.2 Sol and $4/$20 tiers like Claude Opus 5.5—multi-tier routing is clearly moving from an optimization into default architecture.
In practice, though, the cost curve is rarely the part that fails first. The fragile spots show up when a cheap router misclassifies a subtle multi-step tool call and escalates mid-trajectory:
- Context handoff overhead: Replaying 40k+ tokens of tool history into a frontier tier after a Flash model stalls erases a large slice of the per-token savings.
- Routing classifier drift: When upstream APIs or MCP tool schemas change weekly, static routing heuristics quietly degrade before anyone notices.
- Latency jitter: A 90ms router decision looks negligible until it triggers a cold-start fallback across providers.
For teams running two-tier or router-fronted agent pipelines in production right now, where do you draw the escalation boundary—by task type upfront, or dynamically after a failed tool call?
Top comments (0)