Does Anthropic prompt caching actually save money in production?

Yes, dramatically — but only on the right workloads. On our measured production traffic (330 LLM calls across Citare's audit pipeline), Anthropic's native prompt cache hit 85-90% on the cached prefix and delivered 25-35% reduction on total workload cost. The savings come almost entirely from one pattern: long stable system prompts plus short variable user inputs (parsing, classification, structured extraction). On open-ended generation where output dominates the token mix, caching barely moves the needle because you can't cache outputs.

When does prompt caching make sense?

Parsing, classification, structured extraction, and any workload where the system prompt is long (1024+ tokens) and stable across calls while the user input is short and variable. The caching write penalty (~25% cost premium on the first write) gets amortized fast when the cached block is read many times. Below roughly 5 reads per write, caching costs more than it saves. Above 10 reads per write, savings compound rapidly.

When should you NOT use prompt caching?

Open-ended generation where output tokens dominate (creative writing, code generation, long-form content). The output is uncacheable; caching the small input portion barely affects the bill. Also: any workload with constantly changing system prompts (caching has 5-minute ephemeral TTL), or workloads with too few cache hits per write to amortize the 25% write penalty.

Should I build three-layer caching like Prism does?

Only if you serve enough volume to amortize the engineering work. From our Citare data: exact caching paid back below 5% on overall cost (mostly useful for idempotency guarantees, not savings). Semantic caching paid back ~5% and turns ROI-negative below ~10,000 calls per month. Anthropic-native prompt caching paid back 25-35% on the right workloads with zero engineering effort beyond the API parameter. Start with Anthropic native caching alone; layer the others only when volume justifies it.

What's the biggest LLM cost trap that isn't caching?

Failure-retry loops. Burned $20 in one afternoon on Citare's pipeline when a downstream call had an insufficient response token budget, the call failed, the retry logic re-issued the same prompt, the response was again truncated and failed, and the cycle repeated. Caching can't fix this — only proper edge-case provisioning can. Provision response token budgets generously, set hard retry caps, and never let a retry loop run unattended on a prepaid AI key.

Anthropic Prompt Caching: Real Numbers From 330 Production Calls

Real first-party data: Anthropic prompt caching cut Citare's AI bill 25-35% on parsing-heavy workloads. What works, what doesn't, what burned me $20.

By Ravi · May 23, 2026 · Updated May 23, 2026 · 11 min read

anthropicprompt-cachingllm-infrastructurefinopsclaudeprism

I measured Anthropic’s prompt caching on Citare’s real production traffic over 330 LLM calls. Here are the numbers and what they mean for anyone running serious LLM workloads in 2026.

This post is for two audiences: (a) anyone wondering whether prompt caching is worth setting up, and (b) anyone running multi-layer caching and wondering which layer actually pays off. The answer differs sharply by workload type, and most blog posts on this topic don’t have real production numbers to show. I do.

I’m Ravi. I built Prism — an OpenAI-compatible AI gateway that implements three-layer response caching (exact match + semantic + Anthropic-native passthrough). I dogfood it inside Citare, which makes a lot of automated LLM calls to monitor AI-search visibility across five engines. The numbers below come directly from Citare’s logs running through Prism.

TL;DR — the headline numbers

Cache layer	Hit rate	Cost savings on workload	Worth setting up?
Exact (Redis fingerprint)	<5%	1-3%	Idempotency only — not savings
Semantic (Upstash Vector + BGE)	~12% (parse-step only)	~5%	Below 10K calls/month: no. Above: marginal.
Anthropic native prompt cache	85-90%	25-35%	Yes. Day one. No engineering effort.

The headline finding: Anthropic’s built-in prompt caching is, by an order of magnitude, the highest-ROI cache layer for production LLM workloads. Custom application-layer caching is interesting engineering and useful in specific situations, but for raw cost-per-call reduction, the win lives at the provider boundary.

The workload that produced these numbers

The data is from Citare’s audit pipeline — 330 Claude calls across:

10 cold-prospect brand audits (full pipeline: keyword extraction → SERP analysis → AI engine queries → response parsing → scoring)
1 publishable famous-name audit (deeper pipeline with verification steps)
1 Round-2 delta comparison (re-running prior audits and surfacing changes)

The pipeline is heavily parse-and-classify weighted — most calls take a long structured system prompt (the audit framework, response schemas, scoring rubrics) and a short variable user input (the SERP fragment or AI response being analyzed). This is the sweet spot for prompt caching. The numbers would be much worse on a creative writing workload.

I’m flagging that upfront because cache savings are extremely workload-dependent. Your numbers will not match mine unless your workload shape matches.

What each layer actually does and why the numbers differ so sharply

Exact caching (Redis fingerprint) — under 5% hit rate, 1-3% savings

Hash the full request payload (system prompt + user prompt + parameters) → look up in Redis → return the prior response if present.

Why the hit rate is so low: In production, requests are almost never byte-identical. Even when the system prompt is stable, the user input varies. The 5% hits we see come from genuine duplicates — retries of the same request, the same prompt being audited twice, idempotent operations.

What it’s actually good for: idempotency guarantees, not cost savings. If a network blip causes a duplicate request, exact cache returns the prior response and prevents double-charging the user (or running the same expensive analysis twice). That’s a correctness feature, not a cost one.

Verdict: Build it for correctness. Don’t expect it to save money.

Semantic caching (Upstash Vector + BGE-small embeddings) — ~12% hit rate, ~5% savings

Embed the request via BGE-small → cosine-search Upstash Vector → if a prior request scores above a similarity threshold, return its response.

Why the hit rate is higher than exact but still modest: Semantic similarity catches paraphrases and lexical variations of the same intent. In Citare, this hits ~12% in the parse step (different SERP fragments often ask the model to do roughly the same parsing) and almost zero in the audit-generation step (each prospect’s audit is genuinely unique).

The ROI cliff: Semantic caching has real fixed costs — Upstash Vector queries, BGE embedding compute, similarity threshold tuning. Below ~10,000 calls/month, the infrastructure cost exceeds the savings. We’re above that line on Prism workloads, so it pencils. For a smaller operator, it doesn’t.

Verdict: Build it only if you’re at scale AND your workload has high semantic redundancy AND you can afford the engineering cost of tuning the similarity threshold. Otherwise skip it.

Anthropic-native prompt cache — 85-90% hit rate, 25-35% workload savings

Mark a portion of your prompt (typically the system prompt or a long context block) with cache_control: { type: "ephemeral" } in the Anthropic API request. Anthropic caches the cumulative prefix at their inference layer; subsequent calls with the same prefix pay 10% of the normal input cost for those tokens (instead of 100%).

Why the hit rate is so dramatically higher: Because the cache key is the prefix, not the full prompt. A long stable system prompt + varying user inputs hits the same prefix every time. As long as you make calls within the 5-minute ephemeral TTL window, you keep paying 10% on the cached portion.

The 90% / 25-35% math: Anthropic’s pricing makes cached prefix tokens cost 10% of normal (a 90% reduction on the cached portion). But your total workload cost includes uncached input tokens AND output tokens. So a 90% reduction on the cached input translates to 25-35% reduction on total workload cost, depending on the input/output token ratio.

The write penalty caveat: First-time writes to the cache cost 125% of normal input pricing (a 25% premium). If your prefix is rarely reused (≤5 reads per write), prompt caching costs more than not caching. Above ~10 reads per write, savings compound rapidly. Most production workloads sit comfortably above that breakeven.

Verdict: Turn it on. Day one. Zero engineering effort beyond the one cache_control parameter. The savings are real and immediate on the right workloads.

The crucial insight most posts miss: output tokens are the real cost ceiling

Here’s the part that took me a while to internalize, and it changes how you should think about LLM cost optimization:

You cannot cache outputs. Every call’s response tokens are generated fresh and charged at full output pricing.

This means: for workloads where output dominates the token mix (creative writing, code generation, long-form analysis), even perfect input caching barely moves the bill. You can cache the entire 8K-token input down to 10% of cost — but if the response is 6K tokens of generated content at $15/M tokens, the input savings are noise.

Inversely: for workloads where input dominates (parsing, classification, structured extraction with short JSON outputs), input caching is transformational. A 90% reduction on the dominant cost component is a near-90% reduction on the whole bill.

The practical rule: before deciding whether to invest in caching infrastructure, look at your output:input token ratio in real traffic. If output is < 20% of total tokens, caching is high-leverage. If output is > 60%, caching is mostly cosmetic — optimize elsewhere (smaller models, better prompting, output schema constraints).

The sweet spot pattern: long stable system prompt + short variable user input

If you’re designing an LLM workflow from scratch with cost in mind, structure it like this:

System prompt: long, stable, cached. Put your role definition, output schema, examples, scoring rubrics, classification taxonomy — anything that doesn’t change between calls — into the system prompt with cache_control: { type: "ephemeral" }.
User prompt: short, variable. Pass only the input that actually varies (the data to be analyzed, the question to answer).
Output schema: tight. Constrain output format (JSON schema, max length, structured response) to keep output tokens minimal.

This pattern compounds three wins:

Anthropic native cache hits the long prefix at 10% cost
Output tokens stay small so they don’t dominate the bill
The system prompt being stable means cache TTL refreshes naturally as calls continue

This is exactly the pattern Citare’s audit pipeline uses, and it’s why our cache hit rate is so high.

Workloads where caching doesn’t pay off

Three patterns where I’d skip caching investment:

Open-ended generation. Long-form writing, code generation, creative brainstorming — output dominates. Caching the 500-token system prompt does nothing for the 4000-token response.
One-shot pipelines with no reuse. A workflow where each prompt is genuinely unique and the system prompt doesn’t repeat — the 25% write penalty isn’t amortized. You’ll pay more with caching enabled.
Low-volume workloads (under 10K calls/month). The engineering cost of setting up semantic/exact caching layers exceeds the savings. Use Anthropic native caching only (zero engineering cost) and skip the rest.

The $20-in-an-afternoon trap that has nothing to do with caching

I want to end with a war story because it points at the bigger lesson.

I lost $20 in a single afternoon on Citare during early Prism integration. Not to caching costs or runaway tokens — to a retry loop that no one noticed.

The setup: a downstream Anthropic call had a max_tokens budget that was too small for the response the model wanted to generate. The response would get truncated, the JSON parse would fail, the retry handler would re-issue the same prompt with the same (insufficient) token budget, the response would truncate again, parse would fail, retry would fire again. Loop until exhaustion or until I noticed the bill.

Each loop iteration cost a few cents. The loop ran for several hours unsupervised before I caught it. $20 burned.

Anthropic’s prompt cache cannot save you from this. No caching layer can. The fix is correct edge-case provisioning:

Generous response token budgets. When in doubt, set max_tokens 2× what you think you’ll need. The cost of overprovisioning is small; the cost of truncation-into-retry loops is enormous.
Hard retry caps. Never let retry logic run unbounded. Maximum 3 attempts, exponential backoff, then surface the failure to the calling code.
Cost alerts on prepaid API keys. Anthropic, OpenAI, and Google all support spending alerts. Set them. The first warning should fire at 2× your normal daily spend.
Watch your logs. A retry loop announces itself loudly in request logs — same prompt, same timestamps, same parse failure, every few seconds. Surface this pattern.

The bigger thesis here is LLM FinOps — treating AI API spend with the same discipline you’d treat cloud infrastructure spend. Caching is part of it. Failure-mode discipline is the other part. I’m writing that post next.

How to actually set up Anthropic prompt caching (the 30-second version)

For anyone who hasn’t enabled this yet, here’s the minimal change. In your Anthropic SDK call, add cache_control to the system prompt block:

response = client.messages.create(
    model="claude-sonnet-4-6",
    system=[
        {
            "type": "text",
            "text": "<your long stable system prompt here>",
            "cache_control": {"type": "ephemeral"}
        }
    ],
    messages=[
        {"role": "user", "content": "<your variable user input>"}
    ],
    max_tokens=4096
)

That’s the entire integration. No new infrastructure. No engineering effort beyond the parameter. Your next call with the same system prompt hits cache and costs 10% on the cached portion.

Requirements to be aware of:

Cached blocks must be ≥1024 tokens (Sonnet) or ≥2048 tokens (Opus) — smaller blocks aren’t cacheable
Ephemeral TTL is 5 minutes — calls within the window keep the cache warm; longer gaps re-pay the write penalty
Cache key includes model and parameters — switching models invalidates the cache

Bottom line

For most production LLM workloads in 2026, here’s the priority order:

Enable Anthropic native prompt caching. Day one. Single parameter. 25-35% savings on the right workloads.
Structure your prompts to fit the sweet spot. Long stable system prompt + short variable user input + tight output schema.
Skip exact / semantic caching layers until you’re above 10K calls/month and your workload has actual semantic redundancy.
Provision response token budgets generously. Set retry caps. Set cost alerts. Never trust an unattended retry loop.
Read the LLM FinOps post coming next for the broader framework.

Claude Code Review 2026 — From Zero Code to 3 Live SaaS — where the 35-45% cost claim was first cited
How I Run 3 Production AI SaaS on $5/Month of Hosting — the full bootstrapped stack
What is LLM FinOps? (coming next) — the broader framework
How I built Citare in 12 days with Claude Code (coming soon) — the case study showing the audit pipeline these numbers came from

Last updated 2026-05-23. Anthropic’s caching pricing and TTL behavior change occasionally — I refresh this post when material updates ship. If your production numbers look different from mine, tell me on Twitter/X — I’m collecting workload comparisons for a future deep-dive.