Most messages an agent gets aren't hard. "Thanks, that's all." "What's the status of order 4417?" "Find the latest invoice from Acme." A careful, expensive model spends as much answering those as it does on a three-step plan.
Starting today, an agent can have two models. Its main model does the hard work. An optional fast model takes the simple messages, at the fast model's price.
How a message finds its model
Before the agent answers a new message, a small router model reads it and picks one of the two. It reads the newest message and the start of the agent's last answer, because "ok, go ahead" means something different after a plan than after a greeting. Its two options are each model's own description of what it is good at.
- Simple messages, lookups and searches go to the fast model.
- Plans, analysis and long writing go to the main model.
The fast model only starts when the router is confident. When the router isn't sure, takes longer than it should, or the fast model can't handle the conversation (an image it can't see, a chat longer than its context window), the main model answers, exactly as it did before.
| Message | Usually goes to | Why |
|---|---|---|
| "Thanks, that's all" | Fast | Small talk |
| "Find the latest invoice from Acme" | Fast | A lookup |
| "Compare Q3 spend by vendor and draft a note for finance" | Main | Analysis, then long writing |
| "ok, go ahead", right after a plan | Main | The router reads the last answer too |
Either model can hand off
A route is a first guess, not a lock. If the fast model finds a task is bigger than it looked, it hands the conversation to the main model with a short note on what it found so far. The main model can hand a simple follow-up back the same way.
The note travels as a hidden message, so a handoff works between models from different providers, and the person on the other end never hears about it. Each call is told which of the two models it is, so neither one gets confused about its job.
To keep a run from bouncing back and forth, it can hand off to the fast model at most twice. Handing back to the main model is always allowed.
Choosing a fast model
Open an agent's Settings and pick a Fast model under the main one. Flash, Haiku, mini, nano and Lite models all count as fast. Leave it on No fast model and every message goes to the main model, as before.
A fast model needs a slow main model beside it. Two fast models would leave nothing to route between, so that pairing is refused.
What it costs
Each model call is billed to its own model, on the key it runs on. A greeting the fast model answers costs what the fast model charges, and a plan the main model writes costs what it always did. The more of an agent's traffic is simple, the bigger the saving.
Seeing the route
Every step in a trace is tagged fast or slow, and an agent with a fast model gets a Router step at the top of each trace: the model it picked, how sure it was, and why the main model took over when it did. More in See why a message went to the fast model.