What is prompt caching and how much does it save?

Prompt caching stores the processed form of a repeated prompt prefix — a system prompt, document set, or conversation history — so later requests read it instead of reprocessing it. On Anthropic's API a cache read costs 0.1x the base input price, a 90% saving. DeepSeek's V4 cache hits are roughly 97% below base input. The saving applies only to the cached prefix, not to output tokens.

When does prompt caching become worth it?

Writing a 5-minute cache costs 1.25x the base input price on Anthropic's API, so it pays off after a single cache read. A 1-hour cache write costs 2x and pays off after two reads. If you send a large prompt prefix once and never reuse it inside the cache lifetime, caching costs you 25-100% more than not caching.

Why did my prompt cache stop working?

Almost always because the prefix changed. Caching matches on an exact prefix, so a timestamp, session ID, randomized example order, or user name near the top of the prompt invalidates the cache on every request — you then pay write rates forever and never get a hit. Put everything volatile at the end of the prompt, after the cached content.

Quick Answer

Prompt Caching Explained: Cut AI API Costs in 2026

Published: August 17, 2026

The Short Answer

Prompt caching stores the expensive part of your prompt so you stop paying to reprocess it. On Anthropic’s API a cache read costs 0.1x the base input price. On DeepSeek’s V4 models it’s roughly 97% off.

The catch nobody reads: writing the cache costs more than a normal input token. Caching is a bet that you’ll reuse the prefix before it expires. Win the bet and you cut input costs by 90%. Lose it and you paid a 25-100% premium for nothing.

How It Works

Every request re-sends your entire context — system prompt, tool definitions, documents, conversation history. The model reprocesses all of it, every time. If 95% of that is identical across calls, you’re paying repeatedly for identical work.

Caching stores the processed form of a prefix and lets later requests reference it. Three things follow from that, and they explain nearly every caching mistake:

It’s a prefix, not a set. Matching starts at the beginning of the prompt and stops at the first difference.
It’s exact. One changed character invalidates everything after it.
It expires. Typically 5 minutes or 1 hour, refreshed on each hit.

The Pricing Structure

Anthropic’s multipliers, relative to base input price:

Operation	Multiplier	Lifetime
5-minute cache write	1.25x	5 minutes
1-hour cache write	2x	1 hour
Cache read (hit)	0.1x	Refreshes the window

In absolute terms, for Claude Sonnet 5 at $2/MTok base input: writes cost $2.50 (5m) or $4 (1h), reads cost $0.20. For Claude Opus 5 at $5: writes $6.25 or $10, reads $0.50.

Other providers, August 2026:

Provider / model	Base input	Cache hit	Effective discount
Claude Sonnet 5	$2	$0.20	90%
Claude Opus 5	$5	$0.50	90%
DeepSeek V4-Pro (off-peak)	$0.66	$0.022	~97%
DeepSeek V4-Flash (off-peak)	$0.22	$0.007	~97%
GLM-5.3	$1.40	$0.26	~81%
Grok 4.6	$2	$0.50	75%
Muse Spark 1.2	$1.25	$0.15	88%

The Break-Even Math

5-minute cache: write costs 1.25x, read costs 0.1x. Versus 2x uncached calls at 1.0x each:

Cached: 1.25 + 0.1 = 1.35x
Uncached: 1.0 + 1.0 = 2.0x

Profitable after one read.

1-hour cache: write costs 2x.

Two calls cached: 2 + 0.1 = 2.1x vs 2.0x uncached — still losing
Three calls cached: 2 + 0.2 = 2.2x vs 3.0x uncached — now winning

Profitable after two reads. The rule: use 5-minute caching by default; reach for 1-hour only when you’re confident of three or more hits, such as a long agent session or a support conversation.

At ten calls against a cached prefix, the 5-minute cache costs 1.25 + 0.9 = 2.15x versus 10x uncached — a 78% reduction on the prefix.

What This Doesn’t Cover

Output tokens are never cached. If your workload is short prompts and long generations, caching barely moves your bill. Check the split before investing engineering time: caching pays for long-prefix, short-output workloads — RAG over a fixed corpus, agents with large tool definitions, chat over a long system prompt, document Q&A.

It does not pay for one-shot classification with tiny prompts, or creative generation where output dominates.

The Five Mistakes

1. Volatile content at the top. A timestamp, request ID, randomised few-shot order, or user name near the start of the prompt invalidates the cache every single request. You pay write rates forever and never see a hit. Everything that changes goes last, after the cached block.

2. Caching things that aren’t reused. Unique per-request documents in the cached region means paying 1.25-2x for a cache that’s never read. Cache the shared prefix only.

3. Ignoring the minimum. Providers set a minimum cacheable length. Anthropic lowered the minimum to 512 tokens for Opus 5. Below the threshold, caching is silently ignored — you pay full price and see no error.

4. Not measuring hit rate. API responses report cache read and write token counts. If hits aren’t 10x your writes on a steady workload, something upstream is invalidating the prefix. This is the single most useful metric in the whole exercise.

5. Assuming the discount is stable. DeepSeek’s cache-hit price rose up to 12x on August 16, 2026 — from $0.003625 to $0.044 on V4-Pro at peak — even though the percentage discount barely changed, because the base moved. An architecture whose economics depend on one vendor’s unusually generous cache treatment is renting its advantage. See the August 2026 price divergence.

How To Implement It

Restructure the prompt: stable → volatile. System instructions, tool definitions and reference documents first. User input, timestamps and session data last. Do this before touching any cache API.
Cache the largest stable block you have. Usually tool definitions plus reference corpus, often 80%+ of input volume.
Start with automatic caching if offered. Anthropic’s automatic mode takes a single top-level cache_control field and manages breakpoints as conversations grow. Move to explicit breakpoints only when you need finer control.
Default to the 5-minute window. Upgrade to 1-hour only where you can show three or more hits.
Log cache_read_input_tokens and cache_creation_input_tokens on every call, and alert when the ratio drops.
Stack the discounts. Caching multipliers combine with batch discounts and off-peak scheduling. Together they routinely beat switching to a cheaper, weaker model — without the quality loss.

Last verified: August 17, 2026. Rates from official vendor pricing pages.