How does LLM prompt caching actually save money?

When processing a prompt, a transformer model spends the majority of its compute calculating Key-Value (KV) attention matrices for input tokens. If your application sends the same system instructions, documentation, or conversation history repeatedly, the provider stores these precomputed KV tensors in fast GPU memory. Subsequent queries that match the cached prefix bypass redundant attention compute, and providers pass this efficiency back to you as an 80% to 90% price discount.

What are the prompt caching discounts across major AI providers in 2026?

(1) Anthropic Claude 3.7 Sonnet: $3.00/M standard input drops to $0.30/M cache read (90% off); (2) OpenAI GPT-4o: $2.50/M standard drops to $1.25/M cache read (50% off); (3) DeepSeek V3: $0.14/M standard drops to $0.014/M cache read (90% off); (4) DeepSeek R1: $0.55/M drops to $0.14/M cache read (75% off).

What is the minimum prompt size required for caching?

Anthropic requires a minimum prefix of 1,024 tokens for Claude 3.7 Sonnet and 2,048 tokens for Claude 3.5 Haiku. OpenAI automatically caches any prefix longer than 1,024 tokens. DeepSeek automatically caches prefixes longer than 1,024 tokens.

Does prompt caching also reduce Time to First Token (TTFT) latency?

Yes! Because the provider does not have to recalculate attention over thousands of prefix tokens, TTFT is frequently reduced by 50% to 85%. For large 50,000-token context windows, TTFT drops from 4.2 seconds down to less than 700 milliseconds.

💎 90% Cost Reduction Engine

Prompt Caching ROI & Savings Calculator

Calculate exact bottom-line dollars saved with Anthropic, OpenAI, and DeepSeek prompt caching. Model cache hit rates, prefix sizes, and annual gross margin expansion.

⚙️ Prefix & Query Parameters

Interactive model

Reusable Context / Prefix Size 8,000 tokens

1,024 min (Prompt) 25,000 (RAG Docs) 100,000 (Full Codebase)

Monthly API Query Volume 50,000 requests/mo

Cache Hit Rate % (Prefix Reuse) 80% Hit Rate

Average RAG or agent workloads achieve 70% to 85% prefix cache hit rates within the 5-minute TTL.

LLM Endpoint

💵 Financial Savings & ROI

NET MARGIN EXPANSION

Spend Without Caching

$1,425 / mo

Full input rate billed

Spend With Prompt Caching

$475 / mo

Discounted KV-read rate

Net Monthly Cash Saved

+$950 / mo

66.7% cost reduction

Annualized Cash Kept

+$11,400 / yr

Direct profit addition

⚡ Latency (TTFT) Acceleration:

Prompt caching avoids recalculating attention over 8,000 tokens. Time to First Token (TTFT) drops by approximately 68% on cached hits!

Monthly Input Tokens Processed: 400M tokens

Cached Tokens Billed at Discount: 320M tokens

Blended Input Cost per 1M: $0.84 / 1M

Model Impact on SaaS Gross Margins →

Provider Prompt Caching Rules & Discounts (2026 Reference)

Provider & Model	Standard Input	Cache Write Rate	Cache Read Rate	Effective Discount	Minimum Cache Size	Cache TTL Window
Anthropic Claude 3.7 Sonnet	$3.00 / 1M	$3.75 / 1M	$0.30 / 1M	90.0% Off	1,024 tokens	5 minutes (refreshed on hit)
OpenAI GPT-4o	$2.50 / 1M	$2.50 / 1M	$1.25 / 1M	50.0% Off	1,024 tokens	Automatic (typically 5-10 min)
Anthropic Claude 3.5 Haiku	$0.80 / 1M	$1.00 / 1M	$0.08 / 1M	90.0% Off	2,048 tokens	5 minutes (refreshed on hit)
DeepSeek V3	$0.14 / 1M	$0.14 / 1M	$0.014 / 1M	90.0% Off	1,024 tokens	Automatic persistent caching
DeepSeek R1 Reasoning	$0.55 / 1M	$0.55 / 1M	$0.14 / 1M	74.5% Off	1,024 tokens	Automatic persistent caching

Engineering Best Practices for 90%+ Cache Hit Rates

To maximize prompt caching ROI, you must structure your prompt payloads with strict determinism. AI providers match prefixes sequentially from the very first token. If a single dynamic character (such as an unpredictable timestamp or user ID) is placed near the top of the prompt, the entire remaining context cache is invalidated.

// RECOMMENDED PROMPT STRUCTURE:
1. Static System Instructions & Formatting Guidelines (NEVER CHANGES)
2. Tool Schema Definitions & Few-Shot Examples (NEVER CHANGES)
3. Static Documentation / RAG Knowledge Context (CACHE BREAKPOINT HERE)
4. Dynamic User Query & Ephemeral Session Data (ALWAYS AT THE END)

By placing all static tokens at the beginning of the message array and placing dynamic user inputs strictly at the end, your application guarantees that the first 5,000 to 50,000 tokens achieve an unbroken cache hit on every single invocation.

Frequently Asked Questions: Prompt Caching Economics

What is the cost of writing to the cache on Anthropic?

On Anthropic, writing an initial prefix to cache incurs a 25% surcharge over standard rates ($3.75/M on Claude 3.7 Sonnet). However, because subsequent hits cost only $0.30/M, you break even and turn profitable after just two cache reads.

Does OpenAI charge a cache write fee?

No! OpenAI does not charge any cache write premium. Uncached calls pay the standard rate ($2.50/M on GPT-4o), and automatic cache hits receive a flat 50% discount ($1.25/M).

What happens when the 5-minute TTL expires on Anthropic?

If no query touches the cached prefix for 5 minutes, Anthropic evicts the tensors from GPU memory. The next query must write the cache again. However, every time a query hits the cache, the 5-minute TTL timer resets, keeping popular prefixes warm indefinitely in high-traffic applications.