How Prompt Caching Cuts Coding Agent Bills by 85%

Sliding TTL Dynamics & Prefix Invalidation Mechanics

Coding agents (such as Aider, Cline, Continue, and Cursor) send large static system instructions, repository maps (ctags / AST summaries), and tool definitions alongside each iterative prompt. Without prompt caching, a 20-turn agent session that sends an 80k-token repository map reprocesses 1.6 million input tokens, costing $4.80 per run on Claude 3.5 Sonnet. With Anthropic Prompt Caching ($0.30/M cached input vs $3.00/M standard input), that identical session costs only $0.66—an immediate 86.2% cost reduction.

However, prompt caching relies on a 5-minute sliding Time-To-Live (TTL). If a developer pauses to test code locally for more than 300 seconds, the cached KV activations are purged from provider memory, incurring an expensive cache write fee ($3.75/M on Anthropic) on the next turn.

Monthly Operating Cost Matrix: Coding Agent Prompt Caching (100 Devs)

Anthropic Ephemeral Cache Control Injection Script

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

const anthropic = new Anthropic();

// Structure prompt into strictly cached prefixes followed by dynamic user queries
export async function executeAgentTurn(staticRepoMap: string, toolDefinitions: any[], dynamicUserPrompt: string) {
  const response = await anthropic.messages.create({
    model: 'claude-3-5-sonnet-20241022',
    max_tokens: 4096,
    system: [
      {
        type: 'text',
        text: 'You are an autonomous engineering agent with full codebase access.',
        cache_control: { type: 'ephemeral' } // Breakpoint 1: Base Instructions
      },
      {
        type: 'text',
        text: `CODEBASE REPO MAP:\n${staticRepoMap}`,
        cache_control: { type: 'ephemeral' } // Breakpoint 2: Large static repo context
      }
    ],
    messages: [
      { role: 'user', content: dynamicUserPrompt }
    ]
  });

  const usage = response.usage as any;
  console.log(`Cache Read Tokens: ${usage.cache_read_input_tokens || 0}`);
  console.log(`Cache Creation Tokens: ${usage.cache_creation_input_tokens || 0}`);
  return response;
}

Cache Preservation Rules & Failure Modes

• Prefix Stability Invariant: The prompt prefix must be character-for-character identical across turns. Inserting dynamic elements (such as `Current Time: 14:02:11` or dynamic process IDs) ahead of the repo map forces full cache invalidation.
• Background Heartbeat Pings: If an automated build or test suite runs for longer than 4 minutes, the client should trigger a 1-token dummy prompt with the cached prefix to reset the Anthropic 5-minute TTL timer before the real response arrives.

Engineering Q&A: Coding Agent Caching