LLM Cost Optimization: 9 Tactics That Actually Cut Your AI Bill

The nine tactics

• 1. Right-size the model. Route classification, extraction, and formatting to small models; reserve frontier models for reasoning. Biggest single lever.
• 2. Cache prompts. Long system prompts and shared context belong in provider prompt caches at fractional rates.
• 3. Cache responses. Identical requests (same doc, same question) shouldn't hit the API twice. A content-hash cache is an afternoon of work.
• 4. Trim conversation replay. Summarize turns older than N; stop paying for turn one on every turn.
• 5. Cap max output tokens. Output tokens cost multiples of input. Set limits per task instead of inheriting defaults.
• 6. Batch where offered. Batch APIs run at deep discounts for anything that can wait minutes instead of milliseconds.
• 7. Kill silent retries. Wrap retries with budgets and jitter. Retry storms are the classic surprise-invoice story.
• 8. Compress prompts. Delete boilerplate, dedupe instructions, move examples to cached context. 20% fewer input tokens is 20% off that line.
• 9. Put a budget on every model. Optimization without measurement regresses. Per-model, per-feature budgets with alerts keep wins won.

Measure first, then cut

Tactics 1–8 need tactic 9's data to prioritize. AICosts.ai gives you the unified view spend by provider, model, and day, with budgets and exports so you cut the line that's actually biggest, not the one that's easiest to see.

Find your biggest line item →