The article argues that the common practice of pinging cache keepalive every 30 seconds costs 8× more than necessary, with the optimal interval being around 4 minutes. It provides measured economics across Anthropic, OpenAI, Gemini, and DeepSeek, establishing a rule: keepalive only pays when the agent's pause falls between the provider's eviction point and break-even horizon.
Background
Provider prompt caching is a widely adopted cost-saving technique for LLM APIs, but agentic workloads with long pauses between requests often lose cached prefixes. Cache keepalive—periodically pinging to refresh TTL—is a known defense, but optimal intervals have been based on folklore rather than measurement.
- Source
- Lobsters
- Published
- Aug 6, 2026 at 02:52 AM
- Score
- 7.0 / 10