Token levers

​

The question

Our agent is consuming too many tokens. How would you reduce it?

​

The answer

There are several levers. Model routing is one strategy.

• Prune / compact — old context
• Retrieve better — read less
• Cache — reusable context
• Reduce outputs — dedupe logs
• Subagents — isolate exploration
• Model routing — match cost to reasoning
• 90% — max cost reduction from prompt caching — Anthropic’s own published figure
• 85% — max latency reduction on long, cached prompts
• 1 of 6 — caching is one lever — routing is the one this deck traces next

​

Source

Anthropic docs — prompt caching — Only the cache lever gets numbers here because it’s the one with an official published figure. The other five are real, common patterns — no invented percentages attached.

​

Continue the walkthrough