Last updated: October 2026
- Claude charges extra to write a cache and very little to read it: 1.25x base input for a 5-minute write, 2x for a 1-hour write, and 0.05x to read on Claude Sonnet 5.5 and Opus 5.5.
- OpenAI caches automatically from 1,024 tokens on GPT-5.6 and later, and bills cached tokens at 0.1x of uncached input.
- Gemini 2.5 and newer cache implicitly, with minimums of 2,048 tokens (2.5 Flash and Pro) or 4,096 tokens (3.1 Pro Preview).
- All three miss silently when the prefix is too short or changes between calls, so log the cached-token field on every request.
An SEO content pipeline sends the same 5,000 to 20,000 tokens with every request: the style guide, the banned-word list, the entity map, the schema template.
Prompt caching stops you paying full price for that repeated block. The three big providers implement it in three different ways.
How Does Prompt Caching Work on Claude, OpenAI and Gemini?
On every provider the cache keys on a prefix: the start of the prompt, byte for byte. If the first N tokens match an earlier request, those tokens are not reprocessed at full price.
Anthropic makes you opt in. You add a cache_control block, either on the whole request or on up to four explicit breakpoints, per Anthropic’s prompt caching documentation.
OpenAI does it for you. OpenAI’s prompt caching guide states that caching is enabled by default for supported models and a cached prefix stays eligible for 30 minutes after its last use on GPT-5.6 and later.
Google splits it in two. Implicit caching is the default on Gemini 2.5 and newer. Explicit caching lives in the generateContent API and not in the Interactions API, per Google’s Gemini context caching documentation.
Put everything static first (style guide, entity list, schema template) and everything per-article last (title, keyword, SERP notes). A single changed token near the top invalidates the whole prefix.
What Do the Cache Multipliers Actually Look Like Side by Side?
The table uses the multipliers published in each vendor’s docs as of October 2026. It deliberately shows no dollar prices, because base prices change and you should read them from the live pricing page.
| Factor | Claude Sonnet 5.5 / Opus 5.5 | OpenAI GPT-5.6+ | Gemini 2.5 / 3.1 Pro |
|---|---|---|---|
| Opt-in needed | Yes, cache_control | No, automatic | No for implicit; yes for explicit |
| Write cost | 1.25x (5 min) or 2x (1 hour) | No separate write fee stated | Not stated in the docs page |
| Read cost | 0.05x of base input | 0.1x of uncached input | Savings passed on, no figure on the page |
| Minimum prefix | 512 tokens | 1,024 tokens | 2,048 (2.5 Flash/Pro); 4,096 (3.1 Pro Preview) |
| Lifetime | 5 minutes or 1 hour | 30 minutes after last use | Not stated on the page |
One footnote on OpenAI: GPT-6.1 Sol is listed at 0.05x for cached tokens. Earlier OpenAI models use model-specific cached rates, so check the model you actually call.
Is Claude’s Write Premium Worth Paying for a 40-Article Batch?
Yes, once the shared prefix is read more than about twice within the cache lifetime. Here is the arithmetic, with P as the base input price per token and a 6,000-token style guide plus entity map.
Without caching, 40 articles cost 40 x 6,000 x P = 240,000P for that block. With a 5-minute cache on Claude Sonnet 5.5, one write costs 6,000 x 1.25P = 7,500P.
The other 39 reads cost 39 x 6,000 x 0.05P = 11,700P. Total: 19,200P, which is 8% of the uncached figure. This is a calculation from published multipliers, not a measured bill.
It only holds if the 40 calls land within the TTL of each other. Every successful read refreshes a 5-minute entry, but a batch that stalls for six minutes pays the 1.25x write again.
A 1-hour write costs 2x, not 1.25x. If your batch finishes in ten minutes, the extra 0.75x buys nothing. Use the 1-hour TTL only for pipelines that trickle one article per hour or so.
Why Does My Cache Never Hit, and How Do I Catch It?
The failure is silent. Anthropic’s prompt caching documentation says prompts shorter than the minimum are simply not cached and no error is returned. Your bill rises and nothing in your logs complains.
The minimum also varies by model. Claude Sonnet 4.6 needs 1,024 tokens, Claude Haiku 4.5 needs 4,096, and Claude Sonnet 5.5 needs 512. Switching a cheap drafting step from Sonnet to Haiku can quietly turn caching off.
The fix is a one-line check per call. Read cache_read_input_tokens and cache_creation_input_tokens from the Claude usage block, or the cached-token count from the OpenAI and Gemini responses.
Fail the run if call number two in a batch reports zero cached tokens. Call one is allowed to miss; call two missing means the prefix is unstable or too short.
Which Setup Fits an n8n or WordPress Content Pipeline?
Match the provider to how your jobs arrive. A scheduled batch that fires 20 to 50 calls in a few minutes suits Claude’s 5-minute cache, because the write premium is paid once and reads dominate.
A trickle pipeline, such as an n8n workflow that drafts one WordPress post a day, will miss a 5-minute cache every time. OpenAI’s 30-minute window is also too short there, so a daily single-call job should not count on caching at all.
Gemini’s implicit caching asks the least of you. Keep prompts above the model minimum, keep the prefix stable, and read the cache-hit field to see whether it fired.
Whichever you use, keep the rules in one versioned prefix. Our guide to the AI content pipeline with n8n, Claude and WordPress shows where that shared prompt sits in the workflow.
Cache read tokens are priced at a fraction of the base input price, with the lowest multipliers on the newest Claude models.
– Per Anthropic’s published prompt caching documentation
Claude rewards batch jobs with the lowest read price and charges a write premium. OpenAI asks for nothing but a stable prefix. Gemini is automatic on 2.5 and newer.
On every provider, the real risk is a silent miss. Log the cached-token field and alert on zero.
Frequently Asked Questions
Does prompt caching change the output quality?
No. Caching reuses the processed prefix of your prompt, so the model sees the same tokens. Only cost and latency change.
Why is my Claude cache_read_input_tokens always zero?
Most often the cached prefix is shorter than the model’s minimum, or something before the breakpoint changes on every call (a timestamp, a per-article ID). Per Anthropic’s prompt caching documentation, prompts under the minimum are processed without caching and no error is returned.
Do I need to change code for OpenAI caching?
No. OpenAI’s prompt caching guide says caching is enabled by default for supported models. The work is ordering the prompt so the static part comes first.
Is Gemini implicit caching the same as explicit caching?
No. Per Google’s Gemini context caching documentation, implicit caching is on by default for Gemini 2.5 and newer, with savings passed on automatically. Explicit caching is the version where you create and reference a cache yourself.
Which provider is cheapest for a 40-article batch with one shared style guide?
Compare the cache read multiplier first, then the write cost. Claude Sonnet 5.5 reads at 0.05x of base input; OpenAI’s GPT-5.6 and later read at 0.1x. Which is cheaper in dollars still depends on each model’s base input price, which you should read from the live pricing page.
DesignCopy Editorial Team
English 
한국어 