Trends
Two years ago, 140 people a month searched for "prompt cache" in the US. As of July 2026 it is 1,900, up 1,257% over two years. That part is an ordinary technical adoption curve.
Here is the part that is not. The term has been stuck at exactly 1,900 for four consecutive months, and the three-month growth column reads 0%. Meanwhile the term "token costs" is up 1,011% in the same year. People are searching their bill far faster than they are searching the fix, and prompt caching has quietly stopped being a topic because it became a default. If you run anything agentic, it is now the largest single lever on what you pay.
Key takeaways:
- "prompt cache" grew from 140 to 1,900 US monthly searches in two years, up 90% year over year, classified as RAPID in the Rising Trends database (data as of July 2026).
- It has held at 1,900 for four months straight. The three-month growth column is 0%.
- Cache reads cost 0.1x the base input rate on both Anthropic and OpenAI. That is the 90% discount everyone quotes.
- Writing a prefix costs 1.25x. OpenAI's own worked example: across ten requests, one write and nine reads cost 2.15x instead of 10x.
- Anthropic now prices cache reads on Fable 5.1 at 0.025x, a quarter of the standard multiplier, which means the most expensive model has the cheapest cache read.
- The idea is from a 2023 paper that reported time-to-first-token improvements of 8x on GPU and 60x on CPU.
Let's get into it.
The curve
Here is the monthly search volume for "prompt cache" from our database. Read the right-hand side carefully, because the flat top is the finding.
The breakout month is August 2024, when the major providers started shipping caching as a billable API feature. Then a quiet year, a dip to 720 in December 2025, and a step change through early 2026: 880 in January, 1,300 in February, 1,600 in March, 1,900 in April. It has not moved since.
A plateau in a technical term is not a stall. It is usually the moment a technique stops being something you read about and becomes something your SDK does for you. Both major providers now offer automatic caching that places the breakpoint for you.
The company it keeps tells you what is actually on developers' minds.
kv cache, the underlying mechanism, draws 5,400 a month as of July 2026 and is the larger term. Prompt caching is the productised version of it. But look at which of these are moving.
token costs is up 1,011% year over year and inference cost is up 255%. "Context engineering", the 2025 phrase for the same general problem, is down 76%. The vocabulary has moved from how to arrange a prompt to what the prompt costs, which is what happens when a technology stops being experimental and starts appearing on a finance review.
What prompt caching actually is
When a model processes your prompt, it computes an internal representation of every token, called the attention state. Send a near-identical prompt again and it redoes all of that work from scratch.
Prompt caching stores the computed state for a prefix and resumes from it. Anthropic's prompt caching documentation describes it as "resuming from specific prefixes in your prompts", and that word prefix is the important one. Caching only works from the start of the prompt forward. Anything that changes early invalidates everything after it.
The idea is not new. It comes from a paper posted in November 2023, Prompt Cache: Modular Attention Reuse for Low-Latency Inference, later presented at MLSys 2024. The authors precomputed attention states for text segments that recur constantly (system messages, prompt templates, attached documents) and reused them. Their reported improvement in time to first token ranged from 8x on GPU-based inference to 60x on CPU, with no change to model weights.
That is the whole trick. Three years later it is a billing line.
What it costs, precisely
Here is the part most explainers skip. Caching is not free, and the write costs more than a normal token.
| Operation | Multiplier on base input rate |
|---|---|
| Cache write, 5-minute lifetime | 1.25x |
| Cache write, 1-hour lifetime | 2x |
| Cache read | 0.1x |
| Cache read, Claude Fable 5.1 and Mythos 5.1 | 0.025x |
In dollars, from Anthropic's own table: Claude Sonnet 5 is $2 per million input tokens and $0.20 per million cache read tokens. Haiku 4.5 is $1 and $0.10. Opus 4.6 is $5 and $0.50.
OpenAI publishes the same shape and does the arithmetic for you. Its prompt caching guide says cached input is "discounted up to 90%", and gives the break-even directly: writing a prefix once and reusing it once costs 1.35x what that text would cost uncached, against 2x for processing it twice. Over ten requests, one write and nine reads cost 2.15x instead of 10x.
That is the number to take to a budget meeting. Not a 90% discount on your bill, a roughly 79% reduction on the repeated portion of it, once you have paid the write.
The exception is worth flagging because it inverts the usual advice. Anthropic's footnote puts cache reads on Fable 5.1 and Mythos 5.1 at 0.025x rather than 0.1x. The most capable and most expensive model in the lineup now has the cheapest cache read multiplier of any model published. For a long-running agent that reads the same 200,000-token context on every turn, that footnote can matter more than the headline price.
Why agents changed the maths
A chat app sends a short prompt and gets a short answer. An agent sends the entire accumulated context on every single turn.
The Efficient NLP channel frames it exactly right in the title: agents recompute the same prompt. A twenty-turn agent loop with a stable 50,000-token system prompt and toolset pays for that prefix twenty times without caching. With caching it pays 1.25x once and 0.1x nineteen times. That is the difference between an agent product with a margin and one without.
It is also why the growth in our data sits on the cost terms rather than the technique term. The agent platforms we track all hit the same wall at roughly the same time, and caching is the first thing anyone reaches for.
The failure modes nobody warns you about
Caching fails silently. You do not get an error, you get a bill.
You are under the minimum. Prefixes shorter than a model-specific threshold are simply not cached. Anthropic's minimums run from 512 tokens on Fable 5.1, Mythos 5.1 and Opus 5 up to 4,096 on Haiku 4.5 and Opus 4.6. Amazon's Bedrock documentation is blunt about the consequence: place a checkpoint before the minimum is met and "your inference still succeeds, but your prefix isn't cached."
You are in the cost trap. OpenAI's guide documents a range where a prompt sits just below the threshold and costs more than one deliberately padded up to it. Adding tokens can make a request cheaper, which is counterintuitive enough that most teams never test it.
Something tiny changed at the front. Per Anthropic's troubleshooting section, the cache breaks if you change tool_choice, add or remove an image anywhere in the prompt, alter the thinking configuration or effort setting, or put a timestamp inside the cached block. The nastiest one is language-level: tool_use blocks whose JSON keys get reordered, which the docs note happens in Swift and Go, will miss every time while looking correct.
The clock ran out. The default lifetime is five minutes, refreshed on each use. A user who steps away for lunch comes back to a cold cache and a full-price turn.
What it means for anyone buying AI
Cache hit rate belongs on the dashboard. It is the closest thing the AI stack has to a gross margin lever, and it is measurable per request. Teams that track it find prompt ordering bugs they would never have found otherwise.
Prompt architecture is now cost architecture. Stable content goes first: system instructions, tool definitions, reference documents. Volatile content goes last: the user's message, the timestamp, the retrieved chunk. That single ordering rule is worth more than most model-selection decisions.
Cheaper model is no longer the automatic answer. With a 0.025x read multiplier on the top tier and a 4,096-token minimum on the cheapest one, the small model is not always the cheap model for a long-context workload. Run the arithmetic on your actual prefix length before switching down.
What to watch
Whether the term stays flat. Four months at 1,900 says the technique has settled. If it moves again it will be because pricing changed, not because the idea got more popular. The live prompt cache trend page is where that shows up.
Whether the 0.025x multiplier spreads. One provider has broken the 0.1x convention on its flagship. If competitors match it on their top models, long-context agents get materially cheaper and the "use a small model" default weakens further.
Whether caching and retention collide. Caching stores computed state on the provider's infrastructure, which is a direct tension with the zero-data-retention policies enterprises are now asking for. Anthropic's own docs link the two topics from the prompt caching page. That is a procurement fight waiting to happen, and "token costs" rising 1,011% in a year suggests finance is already at the table.
The honest summary is that prompt caching is the least glamorous and most valuable thing in the AI cost stack. It is a flat search curve attached to an enormous number, which is exactly what mature infrastructure looks like.
Want to catch a shift while it is still a small search curve? Read our guide on how to identify market trends, follow the live prompt cache dashboard page, or see what is breaking out right now on the Rising Trends dashboard.
