Guides

Prompt caching explained, with savings

Prompt caching bills a repeated prompt prefix at a fraction of the input price. How big the discount is, and whether writing to the cache costs extra, varies by provider. Prices verified 2026-10-18.

How it works

When consecutive requests start with the same tokens (system prompt, tool definitions, a document, earlier conversation turns), the provider can reuse its work on that prefix. Those tokens are billed at the cached-input price instead of the normal input price. Anything after the first changed token is billed normally, so put stable content first and variable content last.

Some models charge a premium to write a prefix into the cache. In our data that is 16 models: Anthropic and OpenAI, at 1.25× the input price for a standard cache write (Anthropic's 1-hour cache costs 2× input). Every other model bills cache writes as normal input. Some providers apply caching automatically; others need it enabled per request, so check the provider documentation linked from each model page.

Cache prices by model

USD per 1M tokens. "Reads to break even": how many times a cached prefix must be reused before caching is cheaper than not caching.

Does the cache-write premium pay off?

For Claude Opus 5.5, sending a prefix uncached twice costs 2 × $4.00; writing it once and reading it once costs $5.00 + $0.20. Caching wins after 1 read with the 5-minute cache, and after 2 reads with the 1-hour cache (write at 2× input). For all 16 models with a write premium, the 5-minute break-even is 1 read.

Chatbot: savings by cache hit rate

Cost per 1,000 replies. Output tokens are never cached, so savings flatten once output dominates the bill.

Savings on a 30-step coding agent run

Sorted by the share of the run cost that caching removes.

USD, standard list prices. Tags mark announced price changes and temporary sale prices.

Price is not quality

This page ranks by cost only. Models differ in accuracy, instruction following, tool use, latency, rate limits and data policies, and those differences often matter more than the price gap. Shortlist two or three models from the table, run them on a sample of your real inputs, and compare cost per successful result.

Related comparisons

• Gemini 4 Argon vs DeepSeek V4 Pro

More guides

• Cheapest LLM for coding agents
• Cheapest LLM for RAG chatbots
• Cheapest LLM for batch classification and extraction
• Long-context document analysis (400K tokens)
• Cheapest LLM for high-volume customer support

FAQ