Does prompt caching actually reduce agentic LLM costs?

"Don't Break the Cache" evaluates prompt caching — a provider feature that reuses previously computed key-value (KV) tensors from attention layers so a repeated prompt prefix is not recomputed — across three providers (OpenAI, Anthropic, Google) and four flagship models (GPT-5.2, Claude Sonnet 4.5, Gemini 2.5 Pro, GPT-4o). The evaluation runs on DeepResearch Bench, a multi-turn agentic benchmark where agents autonomously run web-search tool calls to answer research questions, across "over 500 agent sessions with 10,000-token system prompts." The headline result: "prompt caching reduces API costs by 41-80% and improves time to first token (TTFT) by 13-31% across providers," and these savings hold "across all three caching strategies" tested.

The paper compares four cache conditions, each defined by where a UUID is inserted to deliberately break the cache: a No Cache baseline (UUID at the start, forcing full recomputation — standing in for dynamic content like timestamps injected into a system prompt at inference time); Full Context Caching (no UUIDs, letting the provider's automatic caching run unmanaged); System Prompt Only Caching (UUID after the system prompt, so only the static instructions are cached); and Exclude Tool Results Caching (UUIDs after the system prompt and after every tool result). The paradox it surfaces: "naively enabling full-context caching can paradoxically increase latency," because providers "automatically trigger cache creation when token thresholds are exceeded," which caches dynamic, session-specific tool results that will never be read again — cache-write overhead with no matching cache-read benefit. System Prompt Only Caching "provides the most consistent benefits across both cost and latency dimensions," and "the cache strategy that maximizes cost savings does not always maximize latency improvement."

This sits in productive tension with Do persistent agents really cost less per token?, which observed 82.9% cache-read tokens in one physician-scientist's 115-day agentic workflow and inferred that persistent, cache-dominant agents should be costed per artifact rather than per token. That note is a single observational case study with no controlled comparison; this paper supplies the missing controlled measurement — a specific, provider-compared range (41-80% cost, 13-31% TTFT) and the caveat that caching has to be managed, not merely turned on, or its benefit can invert. Both agree that the cache, not the raw token count, is now the economically load-bearing layer of a persistent agent, a point also implicit in how Can external state caches let models solve harder problems? treats the cache hierarchy as engineered harness infrastructure rather than a transparent provider default.

The evaluation is bounded to one benchmark (DeepResearch Bench), one system-prompt size (10,000 tokens), and research/web-search agentic tasks — it does not establish that the same 41-80%/13-31% ranges hold for coding agents, customer-support agents, or other tool-call profiles, and the excerpt notes "provider-specific strategy discrepancies" from an ablation (500-50,000 token prompts, 3-50 tool calls) without detailing what those discrepancies are. The implication the evidence does support: caching strategy is a parameter to tune per provider and per workload, not a single switch to flip, and a team that enables full-context caching by default without checking its latency effect is working against the paper's own central finding.