Stop paying frontier prices for formatting dates

Most AI bills are upside down. A gateway sends everything — password resets, date formatting, existential contract disputes — to the same frontier model, because routing each request "correctly" would require reading it first, which costs as much as just answering it. So the three-cent question and the three-dollar question ride in the same seat.

A routing decision is one choice question: small, frontier, or human. Answer it in milliseconds for fractions of a cent, and the expensive seat stays empty except when it matters. Teams routinely cut inference spend by half or more with no quality change, because most traffic was never hard.

The architecture

The router sits in front of your model tier like a bouncer with a guest list. "Simple, factual, low-risk, reversible" goes left; "complex, high-stakes, ambiguous, irreversible" goes right; anything legal, medical, safety-critical, or furious goes to a person.

Build it

curl $WF_URL/v1/decide/model_router \
  -H "authorization: Bearer $WF_KEY" -H 'content-type: application/json' -d '{
  "state": {"request": "Format these 12 dates as ISO strings."}
}'

r = decide("model_router", {"request": text})
tier = r["answers"]["route"]["choice"]
if tier == "small":
    answer = local_model(text)
elif tier == "frontier":
    answer = flagship_model(text)
else:
    queue_for_human(text, context=r)

Log the tier against downstream thumbs-up rates for a month. If "small" answers satisfy at the same rate as frontier did, your old bill was a donation.

Pitfalls

• Confidence gates the savings. Low-confidence routes should default up a tier, not down — a frontier call is cheaper than a wrong answer.
• Humans are a tier, not a failure. "Needs a person" is a successful routing decision. Price it that way in your metrics.
• Revisit quarterly. As your small models improve, the frontier-worthy slice shrinks. Move the boundary and pocket the difference.