Use prompt caching to cut agent loop costs significantly
Implement prompt caching to reduce agent loop costs by up to 5.3x and triage agent bills by 88%.
Prompt caching
Up to 5.3x on agent loops, 88% on triage agents.
Prompt caching reduces the cost of resending conversation history. The prefix is billed at a tenth of the input price, and new turns pay a 1.25x cache-write rate for new content. Agent loops read a median 84% of input from the cache.
Prompt caching is the most effective lever for reducing costs in agentic tasks. It significantly cuts expenses by caching conversation history, making subsequent turns much cheaper to process. Implement this before other cost-saving measures.
Agent tasks resend the entire growing conversation with each turn. Caching stores this history, allowing subsequent requests to read from the cache at a lower rate instead of reprocessing the full input.
- Optimizing for cost and intelligence ↗Anthropic · Model Overview · Sep 3, 2026 · paraphrased
First seen on Subvolts Sep 3, 2026. Updated Sep 11, 2026. Spotted a problem? Tell us.
- Optimizing for cost and intelligenceAnthropic · Model Overview · Sep 3, 2026
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- ChecklistPrinciples for writing high-quality AI agent toolsKey principles for creating effective tools that AI agents can use successfully. · 30 min
- Cost tipAdjust cache duration for pauses in conversationsChoose between 5-minute and 1-hour cache durations based on pause frequency and length to optimize costs. · 10 min
- Cost tipUse multi-model strategies for cost and intelligence balanceCombine lower-cost and frontier models to optimize expenses while maintaining high intelligence. · 15 min
- Troubleshooting fixRecovering from agent container failureRe-provision a new container when the old one fails, allowing the harness to retry tool calls. · 10 min
- Troubleshooting fixRecovering from agent harness failureReboot the harness with a session ID and retrieve the event log to resume agent work after a crash. · 10 min
- Safety tipIsolate agent sandbox from credentialsEnsure credentials are never reachable from the sandbox where Claude's generated code runs to prevent prompt injection. · 10 min
Questions people ask
- Which AI tools does this work with?
- Any current model. It was seen with Claude in the sources.
- Where did this come from?
- Anthropic (Model Overview). Subvolts collected, normalized and credited it; the link above opens the original.