How does prompt caching reduce LLM API costs?

Prompt caching allows AI providers to store the Key-Value (KV) cache of static prompt prefixes (such as system instructions, RAG context, or schemas) in GPU VRAM across requests. Instead of reprocessing and billing full input token compute on every call, providers offer discounts between 50% and 90% for cached token reads.

What is the prompt caching pricing for Anthropic Claude 3.5 Sonnet?

Anthropic charges 1.25x the base input price ($3.75 per million tokens) to write the cache with a 5-minute time-to-live (TTL). When a request hits the cache, reads are discounted by 90% ($0.30 per million tokens) compared to standard input pricing ($3.00 per million tokens).

What is the minimum prompt size required for prompt caching?

For Anthropic Claude and OpenAI GPT-4o, the minimum cacheable prefix is 1,024 tokens. For Google Gemini context caching, the minimum is 32,768 tokens. DeepSeek context caching automatically caches prompts starting from 64 tokens.

What is the break-even cache hit rate for prompt caching?

On OpenAI and DeepSeek, where there is no write premium, prompt caching is ROI-positive on the very first cache hit (>0% hit rate). On Anthropic, due to the 1.25x write surcharge, you need a minimum cache hit rate of approximately 22% (or at least 2 hits per cache write) to achieve net financial savings.

Cloud LLM Financial Optimization Engine

Prompt Cache Savings ROI Calculator

Forecast monthly cloud API bill reductions, break-even cache hit rates, and Time to First Token (TTFT) latency improvements across Anthropic, OpenAI, DeepSeek, and Google Gemini.

01. LLM Model & Pricing Preset

Target Model Provider

In: $3.00/M Cache Read: $0.30/M Write: $3.75/M Out: $15/M

02. Token Payload & Traffic Volume

Static / Cacheable Prefix Tokens 12,000 Tokens

1,024 (Min) 12k (RAG / Schemas) 50k (Docs) 100k (Full Repo)

Dynamic User Tokens Uncached query per call

Output Gen Tokens LLM completion tokens

Monthly Request Volume 100,000 Reqs / Mo

5k 100k (Mid-SaaS) 500k 2M (High Enterprise)

Target Cache Hit Rate (%) 80.0% Hit Rate

10% (Cold) 50% 80% (Typical RAG) 98% (High Affinity)

Uncached Monthly Cost

$4,950

Standard input pricing

Cached Monthly Cost

$2,586

With 80% prompt caching

Monthly Net Savings

+$2,364 / mo

47.8% Total Cost Cut

Annualized Enterprise ROI

+$28,368 / yr

12-month budget optimization

Monthly Cloud Bill Comparison Save $2,364 / mo

Uncached Baseline $4,950 / mo

With Prompt Caching $2,586 / mo

Time to First Token (TTFT)

~75% Faster

Pre-computed KV cache bypasses prefill

Break-Even Hit Rate

21.8%

Min hit rate to offset cache writes

Hit Rate Sensitivity Matrix

50% Hit $1,250/mo

75% Hit $2,100/mo

90% Hit $2,700/mo

95% Hit $2,900/mo

Architectural Tip: Static Prefix Alignment

To ensure cache hits, always place constant system prompts, RAG documents, and tool declarations at the very beginning of the message array. Any dynamic content (like timestamps or user queries) placed before the static block will invalidate the entire downstream cache!

// CLOUD PRICING ARCHITECTURE

Frontier LLM Prompt Caching Pricing Comparison

Official provider token discounts, cache write fees, and minimum prefix lengths.

Provider & Model	Base Input	Cache Read (Hit)	Read Discount	Cache Write Fee	Min Prefix
Anthropic Claude 3.5 Sonnet	$3.00 / M	$0.30 / M	90% Off	$3.75 / M (1.25x)	1,024 toks
Anthropic Claude 3.5 Haiku	$0.80 / M	$0.08 / M	90% Off	$1.00 / M (1.25x)	1,024 toks
OpenAI GPT-4o	$2.50 / M	$1.25 / M	50% Off	$0 (Automatic)	1,024 toks
OpenAI GPT-4o-mini	$0.15 / M	$0.075 / M	50% Off	$0 (Automatic)	1,024 toks
DeepSeek-V3	$0.14 / M	$0.014 / M	90% Off	$0 (Native)	64 toks
Google Gemini 2.0 Flash	$0.075 / M	$0.01875 / M	75% Off	$1.00 / M / hr	32,768 toks

// FAQ

Frequently Asked Questions

Mastering prompt caching mechanics, TTFT gains, and API cost reduction.

How does prompt caching reduce LLM API bills? ↓

In standard LLM inference, every request requires the model to recompute the self-attention Key-Value (KV) tensors for the entire input sequence (the "prefill" phase). Prompt caching preserves precomputed KV tensors in fast GPU memory across consecutive calls. Because providers do not have to burn GPU FLOPs recalculating attention matrices for the repeated prefix, they pass savings of 50% to 90% to developers.

Why does Time to First Token (TTFT) improve with prompt caching? ↓

In long-context RAG or document-heavy workflows (e.g., 20,000 to 100,000 tokens), processing the prompt prefill can take anywhere from 1.5 to 5+ seconds before the first completion token is streamed. With prompt caching, the prefill computation is virtually instantaneous because the KV tensors are loaded directly from memory, reducing TTFT by 60% to 85%.

What causes a prompt cache miss? ↓

Cache misses occur when:

Dynamic tokens in prefix: Inserting a timestamp, random UUID, or user query at the top of the prompt shifts all subsequent token positions, breaking the hash match.
Time-to-Live (TTL) expiration: Anthropic maintains prompt caches for 5 minutes (refreshed on each hit). If 5 minutes pass without traffic, the cache expires.
Model version mismatch: Changing model checkpoints or system parameters triggers a re-write.

What is the break-even hit rate on Anthropic Claude? ↓

Anthropic charges a 25% premium to write the cache ($3.75/M vs base $3.00/M on Sonnet). However, cache reads cost only $0.30/M (a 90% discount). Setting up the math: to offset the 25% write premium, you need at least 1 cache hit for every 3 to 4 requests, establishing a break-even hit rate of approximately 21.8%. Any hit rate above 22% produces positive ROI.