Prompt Caching Cost Savings Calculator
Most production prompts start with the same text on every request — the system prompt, the tool definitions, the shared documents. Providers charge roughly 90% less to re-read that repeated prefix from cache. This page calculates what that is worth for your traffic, and lists the caveats that shrink it.
Prompt caching savings calculator
Savings apply to the repeated prefix only — the rest of each request bills normally, and output tokens are never cached. Anthropic also charges a one-off premium to write the cache; see the caveats below.
What cached input tokens are
When consecutive API requests start with an identical prefix, the provider can skip re-processing those tokens — the model's internal state for that prefix is already computed and stored. You still send the full prompt, but the repeated part is billed at the cached input rate instead of the regular input rate. Across the major providers that discount clusters around 90%, and DeepSeek goes further:
| Model | Regular input $/1M | Cached input $/1M | Discount |
|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $0.50 | 90% |
| GPT-5.6 Terra | $2.50 | $0.25 | 90% |
| Claude Sonnet 5 † | $2.00 | $0.20 | 90% |
| Claude Haiku 4.5 | $1.00 | $0.10 | 90% |
| Gemini 3.5 Flash | $1.50 | $0.15 | 90% |
| DeepSeek V4 Flash | $0.14 | $0.0028 | 98% |
† Anthropic's newest models tokenize the same text into ~30% more tokens — the discount percentage is unaffected, but the absolute savings per request are larger than the list price suggests. The full price landscape is on the LLM pricing comparison.
How the savings are calculated
savings/month = repeated tokens × requests/month × (regular input price − cached input price) ÷ 1,000,000
Worked example. A 2,000-token system prompt is re-sent on 100,000 requests a month — 200M repeated tokens. On GPT-5.6 Sol those tokens cost 200 × $5.00 = $1,000.00 at the regular input rate, but 200 × $0.50 = $100.00 as cache reads. Prompt caching saves $900.00 a month — 90% off the repeated part of the bill — before you change the model, the prompt or anything else.
Four places caching pays off
The system prompt. The classic win: a long instruction block that is identical for every user and every request. If it's over the provider's minimum cacheable length, this is free money — the worked example above is exactly this case.
Shared RAG context. When many queries hit the same document set — a product manual, a policy handbook — the shared documents can sit in the cached prefix. Per-query retrieved chunks, by contrast, differ on every request and rarely cache; the RAG cost calculator shows how those two cost lines split.
Tool and function definitions. Agents re-send their full tool schemas with every model call — often thousands of tokens, dozens of times per task. Because the definitions sit at the top of the prompt and never change mid-session, they cache extremely well.
Batch processing. Classifying 50,000 documents with the same instruction prefix means 49,999 cache hits. Combine caching with batch-tier pricing (often another −50%) and bulk jobs can cost a fraction of the naive estimate.
Caveats before you bank the savings
- Provider mechanics differ. OpenAI caches sufficiently long prefixes automatically and doesn't charge for cache writes; Anthropic requires explicit cache breakpoints and bills a ~25% premium to write the cache (5-minute tier); Google and DeepSeek offer automatic and explicit modes of their own.
- Caches are short-lived. Typical lifetime is 5–10 minutes since the last hit. Bursty or low traffic means frequent cold starts, and the first request of every window pays the full input price.
- Minimum token thresholds. Prefixes below roughly 1,024 tokens usually don't cache at all — a short system prompt may see no discount whatsoever.
- The prefix must be byte-identical. A timestamp, session id or user name near the top of the prompt breaks caching for everything after it. Put static content first, variable content last.
- Output tokens are never cached. The discount applies to input only — generation is billed at the full output rate every time, which is why output-heavy workloads gain the least.
- Not every model supports it. The open-weights endpoints in our price feed (Llama 3.3 70B and gpt-oss-120B via Together AI) publish no cached input price.
Frequently asked questions
Related tools and guides
- LLM Pricing Comparison — regular input, output and blended prices for 16 models
- LLM Token Cost Calculator — your full monthly bill, before caching discounts
- RAG Cost Calculator — where caching helps (and doesn't) in retrieval pipelines
- OpenAI, Claude and Gemini API cost calculators — per-provider monthly estimates