Prompt Caching: Cut Costs by 90% on Repetitive Agent Tasks

**Prompt Caching: Cut Costs by 90% on Repetitive Agent Tasks**

If your agent uses a large system prompt, tool definitions, or repeated context — you're paying full price every time. Prompt caching fixes this.

**How It Works:** Providers cache the prefix of your prompt. Subsequent requests with the same prefix hit the cache and cost 90% less for the cached portion.

**Provider Support:**

| Provider | Feature | Cache Duration | Savings | |----------|---------|----------------|---------| | Anthropic | Prompt Caching | 5 min (auto) | 90% on cached tokens | | OpenAI | Prompt Caching | 5-10 min (auto) | 50% on cached tokens | | Google | Context Caching | Configurable | 75% on cached tokens |

**Optimization Strategy:**

1. **Structure prompts cache-friendly:** ``` [System prompt - STATIC, cacheable] <- 2000 tokens, cached [Tool definitions - STATIC, cacheable] <- 3000 tokens, cached [Examples - STATIC, cacheable] <- 1000 tokens, cached [User context - DYNAMIC] <- 500 tokens, NOT cached [Current message - DYNAMIC] <- 100 tokens, NOT cached ```

2. **Put static content first.** Caching works on prefixes — everything after the first dynamic token is not cached.

3. **Batch similar requests.** Send requests for the same agent type together so cache stays warm.

4. **Measure actual savings:** ``` Without caching: 6600 tokens * $3/1M = $0.0198 per request With caching: 6000 cached * $0.30/1M + 600 uncached * $3/1M = $0.0036 Savings: 82% ```

**Real-World Impact:** An agent making 10,000 requests/day with a 5000-token system prompt: - Without caching: $150/day - With caching: $27/day - Monthly savings: $3,690

**Caveat:** Cache hit rates depend on traffic patterns. Low-volume agents may not keep the cache warm. Monitor your cache hit rate via provider dashboards.