Claude Prompt Caching Pricing: Costs vs GPT and Gemini

Claude prompt caching pricing has three parts. A cache write costs 1.25x the input price for the 5-minute cache, or 2x for the 1-hour cache. A cache read costs 2.5% to 10% of the input price, depending on the model. On Claude Fable 5.2, reading a repeated prompt prefix from the cache costs 97.5% less than sending it fresh.

We checked every price in this post against Anthropic's, OpenAI's and Google's own pricing pages and caching guides on 10 Nov 2026.

This guide covers every current Claude model, shows when the cache pays off, and compares Claude's cache prices with GPT and Gemini. If you are new to tokens, start with our explainer on what a token is and how AI pricing works.

How Claude Prompt Caching Works

Every API request sends a prompt, and much of it is often the same each time. Think of system instructions, tool definitions, a policy document or the earlier turns of a chat. Without caching, Claude processes all of it again, and you pay the full input price every time.

Claude prompt caching saves the model's work on that shared start of the prompt, called the prefix. The next request that begins with exactly the same prefix reads it from the cache instead. That is faster, and much cheaper.

You turn caching on yourself. The simplest way is automatic caching: add one cache_control field at the top level of the request. You can also place up to four breakpoints on individual content blocks. The breakpoints themselves cost nothing, and caching works on all active Claude models.

Your bill then splits input tokens into three kinds:

• Uncached input: Billed at the model's normal input price.
• Cache writes: The first time a prefix is stored, at 1.25x the input price for 5 minutes or 2x for 1 hour.
• Cache reads: Every later request that reuses the prefix. This is where the saving is.

Caching only cuts the cost of input you repeat. Output tokens cost the same with or without it.

Claude Prompt Caching Pricing by Model

The table shows Anthropic's prices per million tokens for the current Claude models and the ones they replaced.

The write prices follow one rule on every model. The read price does not. Anthropic charges three different read rates:

• 0.025x the input price: On Claude Fable 5.2 and Claude Fable 5.1.
• 0.05x: On Claude Opus 5.5 and Claude Sonnet 5.5.
• 0.1x: On every other Claude model, including Opus 5, Sonnet 4.6, Haiku 5.5 and Haiku 4.5.

So the newer models cut the cache read harder than the list price. Opus 5.5's input is 20% cheaper than Opus 5's, but its cache read is 60% cheaper: $0.20 against $0.50. Sonnet 5.5 reads at $0.10, a third of Sonnet 4.6's $0.30.

One caveat when you compare generations. Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text, and Sonnet 4.6 uses the older one. For the full price list of every Claude model, see our Claude API pricing page.

Claude Cache TTL: 5 Minutes or 1 Hour?

The Claude cache TTL, or lifetime, is 5 minutes by default. Each time a request reads the cache, the 5 minutes start again at no extra cost. For 2x the input price instead of 1.25x, you can write to a 1-hour cache.

Anthropic's own maths gives the break-even point. With the 5-minute cache, caching pays off after one cache read. With the 1-hour cache, it pays off after two.

Anthropic's advice on which to choose is clear. If you send a prompt more often than every 5 minutes, stay on the 5-minute cache, because it keeps refreshing for free. The 1-hour cache suits:

• Slower reuse: Prompts used less often than every 5 minutes but more often than every hour, such as a side-agent that runs for longer than 5 minutes.
• Chat pauses: A long conversation where the user might not reply within 5 minutes.
• Rate limits: Cache hits are not deducted against your rate limit, so more hits mean more headroom.

Watch the clock, too. Anthropic measures the lifetime from the start of the request, not the end of the response. If a response takes 4 minutes to stream, a follow-up must start within about 1 minute to read the 5-minute cache.

The Minimum Prompt Length for Each Claude Model

Claude prompt caching has a floor: a prefix shorter than the model's minimum is never cached. Anthropic's minimums are:

• 512 tokens: Claude Fable 5.2, Fable 5.1, Opus 5.5, Opus 5, Sonnet 5.5 and Haiku 5.5.
• 1,024 tokens: Claude Opus 4.8, Sonnet 5 and Sonnet 4.6.
• 2,048 tokens: Claude Opus 4.7.
• 4,096 tokens: Claude Opus 4.6 and Haiku 4.5.

A shorter prompt is processed without caching, and the API returns no error. If both cache fields in the response usage are zero, the prompt was not cached. Anthropic says that when a prompt falls just short, growing it to the minimum is often worthwhile.

Worked Example: A 20,000-Token Prefix Sent 1,000 Times

Picture a support assistant with a 20,000-token system prompt of instructions and product documents. It handles 1,000 questions, close enough together that the 5-minute cache never expires.

We count only the shared prefix, because the questions and answers cost the same either way. With caching, the first request writes the prefix to the 5-minute cache and the other 999 read it. Figures are rounded to the nearest cent.

The ranking changes once the cache is warm. Claude Fable 5.2 has twice the input price of Claude Opus 5.5, yet its cached prefix costs about the same: $4.20 against $4.10. Moving from Opus 5 to Opus 5.5 cuts this job from $10.12 to $4.10, far more than the 20% cut in the list price suggests.

The write premium barely matters at this volume. On the 1-hour cache, Sonnet 5.5 would cost $2.08 instead of $2.05, because 999 reads dwarf one write.

Real traffic will not hit the cache every time, so treat the table as the best case.

Work Out Your Own Saving

You can price Claude prompt caching for your own app in four steps:

1. Measure the shared prefix: Count the tokens that stay the same on every request.
2. Price the writes: Multiply the prefix by the write price, once for each time the cache expires and is rebuilt.
3. Price the reads: Multiply the prefix by the read price for every other request.
4. Add the rest: The changing part of each prompt and all output tokens bill at the normal rates.

Compare that total with the prefix at the plain input price. To price your own mix of models and volumes, run your token counts through our AI pricing calculator.

Which Claude Model Should You Cache On?

Anthropic's own advice is to choose Haiku for simple tasks, Sonnet for most production workloads and Opus for the most complex reasoning. Claude prompt caching shifts the sums within that advice:

• Haiku 5.5: The lowest cached price of any Claude model, at $0.01, but only on prompts up to 100,000 tokens. Above that, every price rises fivefold.
• Sonnet 5.5: Reads at $0.10, so a large shared context costs little to resend.
• Opus 5.5 and Fable 5.2: Both read at $0.20. If your top-tier work resends long context, Fable 5.2's higher list price matters less than it looks.

Output still bills at full price, so caching helps most where prompts are long and answers short.

How Claude's Cache Prices Compare With GPT and Gemini

How does Claude prompt caching stack up? OpenAI and Google cache too, on different terms. The table sets similar tiers side by side, per million tokens. OpenAI's rates are for prompts up to 272K tokens.

In the mid and budget tiers shown, Claude and OpenAI list the same input, write and read prices. At the top, they part ways. GPT-6 Astra reads at 0.1x, so the worked example above would cost $20.23 cached on Astra, against $4.20 on Claude Fable 5.2.

OpenAI Prompt Caching: On by Default, Longer Lifetimes

OpenAI prompt caching is on by default for supported models. From GPT-5.6 onwards, OpenAI charges for cache writes: 1.25x the input price for the default 30-minute lifetime, or 2.5x for a 12-hour lifetime in the Responses API.

Reads cost 0.1x the input price on most of these models, and 0.05x on GPT-6.1 Sol and GPT-6.1 Terra. Earlier models such as GPT-5.5 have no cache-write charge. Unlike Claude, cached input tokens still count towards OpenAI's tokens-per-minute limits. See the OpenAI prompt caching guide for the details, and our OpenAI API pricing page for every GPT model.

Gemini Context Caching: Implicit, With No Write Price Listed

Gemini context caching is implicit by default for Gemini 2.5 and newer models. Google passes on the saving automatically when a request hits the cache, and its pricing page lists no write price for implicit caching. Explicit caching, which adds an hourly storage price, is not supported on Gemini 4 Argon and later.

Two things make Gemini's prices less certain. First, implicit hits are not guaranteed. Second, the cached prices are introductory: Gemini 4 Argon's rises from $0.10 to $0.20 on 1 Dec 2026, and Gemini 3.8 Flash's from $0.075 to $0.15 on 1 Jan 2027. Google's context caching guide lists the minimums, and our Gemini API pricing page has every rate.

Prompt Caching vs Batch API: Use Both

The Batch API is Anthropic's other big discount: 50% off input and output for requests you send as an asynchronous batch. Claude Sonnet 5.5 drops to $1/$5 per million input/output tokens, and Claude Haiku 5.5 to $0.05/$0.25 on prompts up to 100K tokens.

On prompt caching vs Batch API, you do not have to choose. Anthropic says its caching multipliers stack with the Batch API discount, so Claude prompt caching works inside a batch too. The two cut different things:

• Prompt caching: Cuts only repeated input, but works on live traffic.
• Batch API: Cuts all input and output by half, but results come back later.

There is one catch. Batch requests can run at the same time and in any order, so Anthropic gives cache hits in a batch on a best-effort basis.

Its own recipe is to send one request with the shared prefix and a 1-hour cache block first. Once that finishes, submit the rest. Anthropic says this usually beats the 5-minute cache, because batches commonly take between 5 minutes and 1 hour.

For more ways to trim a bill, see our AI cost optimisation FAQ.

Five Ways to Raise Your Claude Prompt Caching Hit Rate

The prices above only help if requests actually hit the cache. These habits come straight from Anthropic's prompt caching guide:

1. Break at the last stable block: Put the breakpoint on the last block that stays identical across requests, not on a block that changes.
2. Put changing content last: Timestamps, per-request context and the incoming message go after the breakpoint.
3. Keep settings steady: Changing thinking settings or effort can invalidate cached prefixes, and so can adding images or changing tool use settings.
4. Add a second breakpoint in long chats: The cache lookup checks back 20 blocks. If a growing conversation pushes past that, a second breakpoint keeps the hits coming.
5. Measure it: Track the cache write and cache read token counts in every response, and work out your real hit rate.

The Anthropic prompt caching guide has worked code for each pattern.

Pricing Traps to Watch

A few details of Claude prompt caching pricing can turn an expected saving into a surprise on the invoice:

• Haiku 5.5's 100K threshold: Cache reads and writes count towards the prompt length. A prompt over 100,000 tokens bills the whole request at $0.50/$2.50 per million input/output tokens, with cache reads at $0.05.
• Silent misses below the minimum: A prompt under the minimum is not cached, and no error tells you so.
• US-only inference: On Claude 4.6 and later models, setting inference_geo to US applies a 1.1x multiplier to every token price, cache writes and reads included.
• The streaming clock: A long response eats into the 5-minute lifetime, so slow follow-ups miss the cache.

Claude Prompt Caching: Common Questions

Do Cache Hits Count Towards Claude Rate Limits?

No. Anthropic says cache hits are not deducted against your rate limit. That makes caching useful for throughput as well as cost.

Can Other Customers Read My Cache?

No. Anthropic isolates cache entries between organisations and, on the Claude API, between workspaces within one organisation.

Can I Clear the Cache Myself?

No. There is no way to clear it manually. Cached prefixes expire after at least 5 minutes without use.

Is Claude Prompt Caching Worth It?

For almost any app that resends a long, stable prompt, yes. Claude prompt caching pricing rewards one habit: send the same prefix often, and keep it identical. Do that, and the repeated part of your input costs 2.5% to 10% of the list price.

Start with three checks:

• Is your shared prefix over the minimum? That is 512 tokens on the newest Claude models.
• Does it come back within 5 minutes? If not, the 1-hour cache needs only two reads to pay off.
• Which model reads cheapest for your job? Compare cached prices, not just list prices. For more on Anthropic's Opus flagship, read our Claude Opus 5.5 guide.

For this month's price changes across every vendor, see our AI API pricing November 2026 roundup.