1. Compress your prompts. Trim filler language and redundant instructions. A tightened prompt often keeps the same quality at 30–50% fewer tokens.
2. Enable prompt caching. If your system prompt is long and reused, turn on your provider's native caching. You pay full price once, not on every call — the single highest-leverage fix available.
3. Add semantic caching. Instead of matching identical questions only, embed prompts and reuse responses for semantically similar ones. Production systems often see 60–70% cache hit rates. Note: caching discounts input tokens only, never output.
4. Route by complexity. Send simple queries to a cheap model, reserve premium models for requests that truly need deep reasoning. Often 30%+ of traffic doesn't need the top-tier model at all.
5. Cascade for uncertain cases. Start with the cheapest model; escalate only when its output looks uncertain. Keeps average cost low while preserving a safety net.
6. Constrain output length and format. Structured output (JSON, tables) and explicit length limits cut output tokens — and output cost — with minimal effort.
7. Batch what can be batched. Group tool calls and retrieval queries to amortize overhead, especially in agentic workflows making many small calls.
8. Monitor cost like any other metric. Track cost per request/feature, set budget alerts. Teams that build this feedback loop typically find an extra 10–20% in savings every quarter.