Claude Prompt Cache TTL Burn & Cost Calculator (2026)

Anthropic Claude charges a +25% write penalty ($3.75/M) to create a prompt cache, which expires every 5 minutes of inactivity. Simulate cache evictions, rewrite penalties, and determine if keep-alive pings save you money.

Claude Prompt Caching Architecture: 5-Minute TTL Refresh Loops & Write Penalty Economics

Anthropic's prompt caching on Claude 3.5 Sonnet and Haiku provides a massive 90% discount on cache read tokens ($0.30/1M vs $3.00/1M on Sonnet). However, writing to the cache incurs a 25% premium ($3.75/1M), and cache entries expire after an exact 5-minute time-to-live (TTL) window. If an application sends requests at 6-minute intervals, every single request pays the 25% cache-write penalty without ever reaping the 90% read discount, causing net API expenses to rise by 25% compared to completely un-cached requests.

Production Pipeline: Prompt Cache Heartbeat Daemon in Node.js

For multi-tenant SaaS agents with large static system prompts (20,000+ tokens of API schemas and company policy), running a lightweight background heartbeat every 4 minutes and 30 seconds keeps the KV cache perpetually hot:

import Anthropic from '@anthropic-ai/sdk';

const anthropic = new Anthropic();

export class CacheWarmer {
  private timer: NodeJS.Timeout | null = null;
  private readonly staticSystemPrompt: string;

  constructor(systemPrompt: string) {
    this.staticSystemPrompt = systemPrompt;
  }

  public startHeartbeat(): void {
    // 4 minutes 30 seconds: prevents 5-minute TTL eviction
    this.timer = setInterval(async () => {
      try {
        const response = await anthropic.messages.create({
          model: 'claude-3-5-sonnet-20241022',
          max_tokens: 1, // Minimize output token cost
          system: [
            {
              type: 'text',
              text: this.staticSystemPrompt,
              cache_control: { type: 'ephemeral' }
            }
          ],
          messages: [{ role: 'user', content: 'ping' }]
        });
        console.log(`[CacheWarmer] Cache refreshed: ${response.usage.cache_read_input_tokens} read tokens`);
      } catch (err) {
        console.error('[CacheWarmer] Refresh failed', err);
      }
    }, 270000);
  }

  public stopHeartbeat(): void {
    if (this.timer) clearInterval(this.timer);
  }
}

Production Rules & Cache Breakpoint Guidelines

• Minimum 1,024 Token Threshold: Claude prompt caching does not activate on blocks shorter than 1,024 tokens (2,048 tokens on Haiku). Small prompt prefixes will be billed at standard input rates.
• Maximum 4 Cache Breakpoints: Anthropic allows a maximum of 4 `cache_control: { type: 'ephemeral' }` blocks per API call. Place breakpoints at logical boundaries: static tools schema → company handbook → conversation history.
• Audit Response Headers: Track `cache_creation_input_tokens` vs `cache_read_input_tokens` in CloudWatch/Datadog to alert on cache degradation.

Frequently Asked Claude Cache TTL & Heartbeat Questions

Q: Why is a warm heartbeat ping cost-effective for large Claude system prompts?
A: Writing a 30,000-token system prompt to cache costs $0.1125 on Claude 3.5 Sonnet. Sending a 1-token heartbeat ping every 4m 30s keeps the cache active for $0.009/ping ($1.20/day), ensuring high-frequency user requests always hit the 90% discount.

Q: How many prompt cache breakpoints can be defined in a single Claude request?
A: Anthropic allows up to 4 explicit `cache_control: { type: 'ephemeral' }` blocks per message payload, allowing modular caching of tool schemas, organizational knowledge, and session history.