Intermediate

Best Practices: The Token Optimization Playbook

The whole course as an actionable playbook - a maturity model, a 30-day plan, and the checklist to run before anyone proposes cancelling anything.

The Token Optimization Maturity Model

Level	Description	Typical State
0 - Blind	No token visibility; the bill is a surprise	Most orgs at first invoice shock
1 - Aware	Usage dashboards per feature and team	Knows where the money goes
2 - Optimized	Caching, routing, output control deployed	50-80% below the naïve baseline
3 - Governed	Budgets, alerts, review-time cost checks	Costs stay optimized
4 - Self-sustaining	Chargeback; teams own their bills	Optimization is cultural

The 30-Day Optimization Plan

The Pre-Cancellation Checklist

Before any "cancel the AI subscription / kill the AI feature" decision, verify:

• ☐ Do we know cost per successful task (not per seat / per call)?
• ☐ Has a caching pass been done? (Cache hit rate under 50% on top routes = no.)
• ☐ Is anything still defaulting to the frontier model that doesn't need it?
• ☐ Are outputs tuned, or still model-default verbose?
• ☐ Are agents bounded with per-run token budgets?
• ☐ Is non-interactive volume on batch pricing?

Anti-Patterns to Avoid

• Optimizing before measuring - the week-1 instinct, and the wrong order. You will compress a prompt that accounts for 0.5% of spend while the real 40% sits unmeasured.
• Cost-per-token tunnel vision - cheap models that fail and retry cost more end-to-end. Cost per successful task, always.
• One-time heroics - a cleanup without governance decays within two quarters as new features ship unoptimized.
• Over-compression - squeezing prompts until quality drops, then paying it back in retries and user churn. The goal is efficient, not minimal.
• Blocking adoption to save tokens - the point of optimization is to make MORE AI affordable, not less. Cheap tokens should accelerate adoption, not gate it.

Ten Rules to Remember

1. Output tokens cost 3-5× input - tune the talkative side first.
2. Cache what repeats; structure prompts so more repeats.
3. The smallest model that reliably does the task is the right model.
4. Context is a budget, not a bin.
5. Batch anything a human doesn't wait for.
6. Agents need token budgets like loops need exit conditions.
7. Measure cost per successful task, never per token.
8. Stable prefix, volatile suffix.
9. A cost regression is a bug - review it like one.
10. Optimization's purpose is more AI per dollar, not less AI.

Where to Go Next