It's tempting to wire every feature to the smartest model available and call it done. But model choice is a product decision with three axes-quality, latency, and cost-and most features don't sit at the same point on all three. The job isn't to pick the top of a leaderboard; it's to match the model to what the task actually demands.
Match the tier to the task
Modern model families come in tiers for a reason. A frontier model (Claude Opus, or Fable for the hardest long-horizon work) earns its cost on genuinely hard reasoning-multi-step analysis, agentic loops, the stuff where a wrong answer is expensive. A balanced model (Claude Sonnet) is the workhorse for most product features: fast enough to feel responsive, smart enough for real work. A fast, cheap model (Claude Haiku) is the right call for high-volume, well-scoped jobs-classification, extraction, routing-where latency and cost matter more than deep reasoning.
Don't ask "which model is best?" Ask "what does this task need?" The answer is usually cheaper and faster than the frontier.
Mix tiers in one product
The best systems don't pick one model-they route. A cheap model classifies or triages; a frontier model handles the hard cases it escalates. You get the latency and cost of the small model on the common path and the intelligence of the big one only where it pays off.
How I decide per feature:
- High stakes, hard reasoning → frontier tier. Worth the cost and latency.
- Most interactive features → balanced tier. The default sweet spot.
- High volume, narrow task → fast tier. Classification, extraction, routing.
- Unsure? → start cheaper, measure quality with evals, move up only if you must.