# Can routing a turn to a cheaper model cause a silent failure?

Yes. A router that swaps in a cheaper model for one turn doesn't fail loudly when that model gives a worse answer. The call still returns, the agent still completes the task, and the trace reads as a success exactly like it would with the frontier model, so nothing in the trace points at the routing decision as the cause.

The failure compounds quietly because model choice usually isn't logged as a change worth [tracing a quality drop back to](/answers/cause-attribution/how-do-you-find-which-change-caused-a-quality-drop): a router can switch models turn by turn on its own confidence estimate, with no deploy and no commit marking when a given turn ran on the cheaper model. Without that record, a quality drop noticed well after the fact has no single decision to blame, only a pattern spread across however many turns got routed away from the frontier model.

Pinning one model per workflow step, instead of letting a router pick per turn, keeps that record: a regression is then attributable to one change, model or prompt, not to a routing decision nobody wrote down.

---

Sources:
- Hacker News comment, "Launch HN: Tokenless (YC S26) – Automatic model switching to save money": https://news.ycombinator.com/item?id=49107015 (fetched 2026-10-12)

Source: https://tessary.ai/answers/silent-failures/can-routing-a-turn-to-a-cheaper-model-cause-a-silent-failure
More on Silent failures: https://tessary.ai/answers/silent-failures
From Tessary, agent reliability for AI agents in production: https://tessary.ai