playbook 1 of 10

Model routing

Match each job to the right model and cost tier.

The cheapest way to lose money on AI is to send every request to your most capable model. The second cheapest is to send everything to your weakest one and pay for it in quality. Routing is the discipline of matching each job to the right tier.

The decision

Route on the job, not the user. Three questions decide the tier:

Does this step define output quality, or merely move data?
How often does it run?
What does a wrong answer cost here?

Reasoning-defining, low-volume, high-cost-of-error steps get your strongest model. High-volume, mechanical steps get the cheap, fast one. Most systems are a mix.

Route on the job, not the user

A common mistake is giving "premium" users the expensive model for everything. Instead, route per step: the same user's planning step and their formatting step belong on different tiers.

A worked default

Job	Tier	Why
Planning / synthesis	Strongest	Sets the ceiling on quality
Execution / tool use	Balanced	Good judgment, reasonable cost
Extraction / formatting	Cheapest	High volume, low ambiguity

Make it config, not code

Put the routing table in configuration so a model price change or a new tier is a deploy of data, not a rewrite. Measure cost-per-task and quality-per-task separately; routing is the knob that moves them in opposite directions until you find the balance your product needs.

Key takeaways

Match each step to a tier using: defines-quality, frequency, cost-of-error.
Strong model for reasoning; cheap model for high-volume, low-ambiguity work.
Keep the routing table in config so price changes are a deploy of data.
Track cost-per-task and quality-per-task as separate numbers.