The pitch for routing is easy to make and hard to believe: send each message to the cheapest model that can handle it, and your bill falls by half without anyone noticing. Here is what the arithmetic says, modelled over roughly two million messages at the task mix we expect.
The headline number
Modelled against an unrouted baseline, cost per message falls 54% against the same workspaces' pre-routing baseline. The spread is wide: the tenth percentile saw 21%, the ninetieth saw 71%. What separates them is not sophistication. It is task mix.
Workspaces doing mostly extraction, classification and rewriting save enormously, because that work never needed a frontier model. Workspaces doing mostly open-ended analysis save comparatively little, because their work genuinely does need the expensive tier.
Why the savings are so lopsided
The distribution of task difficulty in real usage is not what people assume. When we classify traffic, roughly 55% of messages fall into classes where a fast-tier model scores within noise of a frontier one. Another 19% is coding, where balanced-tier models are strong. Only about a quarter of traffic genuinely benefits from frontier or reasoning models.
But before routing, that quarter of the traffic was setting the model choice for all of it. People pick a model at the start of a session and keep it. Nobody downgrades to a cheaper model for a quick rewrite, because switching is friction and the saving on any single message is invisible.
Routing is not smarter than a person. It is just willing to make the choice 15,000 times a month, which a person never will.
The quality question
The obvious worry is that we are buying savings with quality. The honest answer is that we cannot know until real traffic runs. The mechanism by which quality would degrade is clear enough to state, though: routing trades a small amount of ceiling for a large amount of cost, and on any given message the cheaper model is occasionally the wrong call.
That is a real cost and we should not wave it away, which is why routing aggressiveness is a setting rather than a default, and why pinning a thread to a specific model is always available. We will publish measured quality deltas once we have them.
What we expect to see
- Latency should improve. Most traffic moves to faster models, so median time to first token falls. We expect people to notice this more than any quality change.
- Feedback gets more useful. When every message goes to the same model, a thumbs-down tells you little. When routing is choosing, a thumbs-down is a signal about a specific model on a specific task class, and it improves the next decision.
- The aggressive setting probably will not be worth it. On the modelled mix it saves another 9%, which is unlikely to pay for the extra misroutes. We will find out.
What we would tell a team starting out
Run without routing for two weeks first. The usage report from that period is the only honest baseline you will get, and it tells you whether your task mix is the kind that benefits. If more than half your messages are drafting, extraction or classification, turn routing on and expect a large number. If you are mostly doing long-document analysis, expect a modest one, and consider whether the frontier tier is where your money should be going anyway.