CachingBeginnerHigh impact
Use prompt caching for repeated system prompts
Long, repeated system or tool-definition blocks can cost 90% less when cached. Restructure prompts so static content comes first.
OpenAIAnthropicGoogleDeepSeek
Savings library
Cut your token spend
A practical, opinionated set of techniques for reducing AI API costs without sacrificing quality.
BatchingBeginnerHigh impact
Route async workloads to Batch APIs
Most providers offer 50% off when results can return within 24 hours. Ideal for evals, backfills, and enrichment.
OpenAIAnthropicGoogle
RoutingIntermediateHigh impact
Route by task complexity, not by default to your largest model
Classify intent with a small model, then escalate. A simple router can cut spend 60%+ on mixed workloads.
Prompt designBeginnerMedium impact
Cap output tokens deliberately
Output tokens are typically 4–5x the price of input. Set max_tokens, request concise schemas, and prefer structured outputs.
ContextIntermediateMedium impact
Prune context aggressively
Drop turn-history older than what the model needs. Summarize older state into a short memory block.
Prompt designBeginnerMedium impact
Use structured outputs over freeform JSON parsing
Native structured outputs reduce retries and wasted tokens from malformed JSON.
OpenAIAnthropicGoogle
Model selectionBeginnerHigh impact
Use a small model for classification and extraction
Sub-$0.30 models handle most extraction and routing tasks at near-frontier accuracy.
OpenAIGoogleDeepSeek
MonitoringIntermediateMedium impact
Instrument per-feature token spend
Tag every call with feature, route, and model. You cannot optimize what you cannot attribute.
EstimationIntermediateMedium impact
Estimate before you ship
Run a 1000-call simulation against representative traffic. Project monthly spend before launch.