Any change to the prefix before the breakpoint, including a timestamp or per-request data. Changing tool definitions or thinking parameters also invalidates it. And time: past the TTL, the entry is gone.

Table of Contents

General 14 min read

Prompt Caching Explained: How a 75% Price Cut Actually Works

Prompt caching explained: Fable 5.1 cut cache reads from $1.00 to $0.25 per million tokens. What caching is, why it works, and how to teach it to kids at home.

Prompt caching explained in one sentence: it saves the model’s already-processed version of the beginning of your prompt, so on the next request that part gets read from storage instead of computed from scratch. On September 1, 2026, Anthropic shipped Claude Fable 5.1 with cache reads priced at $0.25 per million tokens, down from $1.00. Same model family, same base input price of $10 per million, and the repeated part of every request suddenly cost a quarter as much.

That’s not a small accounting detail. For any AI tool your kid uses that carries a long set of instructions or a document across many turns, this is the mechanism that decides whether the tool is affordable.

Key Takeaways

Caching works on prefixes. It stores the processed state of everything up to a marked breakpoint, and reuses it only if that exact prefix appears again, unchanged.
Cache writes cost 1.25x the base input price for a 5-minute lifetime, or 2x for one hour. Cache reads normally cost 0.1x base input, and on Fable 5.1 and Mythos 5.1 they cost 0.025x.
On Fable 5.1, that means cache reads at $0.25 per million against a $10 base input price, the 75% cut announced September 1, 2026.
The default cache lifetime is five minutes, measured from the start of the request that writes or reads it. Long gaps between turns mean the cache is gone.
Caching only pays off if the front of your prompt stays identical. One timestamp in the wrong place and the cache never matches.

What caching actually stores, and why prefixes matter

When a model reads your prompt, it doesn’t just look at the words. It builds an internal representation of them, layer by layer, and that computation is what you pay for as input tokens. Prompt caching stores that intermediate state so the next request can skip rebuilding it.

The constraint is that it only works on prefixes, meaning the front of the prompt, from the beginning up to a marked point. The Anthropic caching documentation describes the mechanism precisely: the system computes a hash of everything up to your designated breakpoint and checks for a matching stored entry. If it finds one, the read is nearly free. If not, it processes the full prompt and caches the prefix.

Why prefixes only? Because a transformer’s representation of token 500 depends on tokens 1 through 499. Change token 3 and everything after it changes too. So the cacheable part has to be the stable beginning, and the changing part has to come last.

The documentation’s guidance is a single sentence worth memorizing: place the cache marker “on the last block that stays identical across requests. If your breakpoint lands on changing content (timestamps, per-request data), the prefix hash will never match previous entries.”

What can be cached: tool definitions, system messages, text messages from both sides, images and documents, and tool results. There’s also a minimum length. On Fable 5.1, Mythos 5.1, Opus 5, Fable 5, and Mythos 5 the minimum is 512 tokens. Older models need 1,024, 2,048, or even 4,096. Shorter prompts simply don’t cache, with no error.

The pricing math, with the Fable 5.1 numbers

Caching has three price tiers, expressed as multipliers of the base input price:

Cache write, 5-minute lifetime: 1.25x base input
Cache write, 1-hour lifetime: 2x base input
Cache read (a hit): 0.1x base input, or 0.025x on Claude Fable 5.1 and Mythos 5.1

Plug in Fable 5.1’s base input price of $10 per million tokens. Cache write: $12.50. One-hour write: $20. Cache read: $0.25. Before the change, cache reads on Fable 5 were $1.00 per million, at the standard 0.1x multiplier. MarkTechPost’s September 1, 2026 writeup reported the drop from “$1.00 to $0.25 per million tokens,” with base input and output rates unchanged at $10 and $50.

Now the break-even, which is the part that makes caching a decision rather than a free lunch. A 5-minute write costs 1.25x. A read costs 0.1x on most models. So you pay a 25% premium once to save 90% on each subsequent read. The pricing documentation states it plainly: caching pays off after one cache read for the 5-minute duration, or after two reads for the 1-hour duration.

On Fable 5.1, with reads at 2.5% of base input, the arithmetic gets lopsided in caching’s favor. Pay 1.25x once, then 0.025x forever after. That’s why this particular price cut mattered enough to lead the release announcement.

The same release reported 52.6% on Terminal-Bench-Science 0.1 against 24.7% for Fable 5 and 29.0% for Opus 5, so the model got better and the repeated-context cost dropped. Those two things together are why the September announcement changed what agentic tools cost to run.

Why this decides whether your kid’s AI tool exists

Here’s the connection to a family. Almost every useful AI product sends a large, unchanging block of text on every single request: system instructions, a tool list, a document, a syllabus, a set of class materials. Then it appends the one new thing the user said.

Consider a homework helper with a 20,000-token instruction set. Without caching, every question costs 20,000 tokens of input plus the question. With caching, the first question costs 20,000 × 1.25 and every subsequent question inside the cache window costs 20,000 × 0.025 on Fable 5.1. At $10 base input, that’s 25 cents once, then two-thirds of a cent per question instead of 20 cents.

That difference is the whole business model for AI tools in schools. Google’s August 2026 expansion of Gemini in Classroom, which grounds student prompts in a teacher’s uploaded class materials, is exactly the pattern caching was built for: a large stable document plus one small changing question. Multiply by a district and it’s the difference between a pilot and a cancellation.

Two practical consequences your kid will actually notice:

Long conversations get cheaper per turn, not more expensive. Every turn resends the whole history as input, which sounds ruinous, but the history is a stable prefix and caches well. The tokens explainer covers why history is resent at all.

The five-minute window is real. The default cache lifetime is five minutes, measured from the start of the request that writes or reads it. If your kid asks a question, goes to dinner, and comes back 20 minutes later, the cache is gone and the next request pays full price. Documentation notes that if a response takes four minutes to stream, follow-ups must start within about a minute to reuse that cache.

How to Teach Your Kid About Prompt Caching

Caching is about doing setup work once. Every kid already understands that intuitively; the job is naming it.

Ages 5–8: Set the table once

Before a meal, have your kid help set the table: plates, forks, cups. Then eat one course. Then ask: “Do we have to set the table again for dessert?” No, because the setup is still there. Then clear everything and ask again. Name it: “Computers do the same thing. They set up the same information once, keep it for a few minutes, and then it goes away and they have to set it up again.” Five minutes later, ask if the table is still set. That’s the cache expiring.

Ages 9–12: The math-facts warm-up

Give your kid a page of problems that all start the same way, like “Convert to a fraction, then add 3/8: …” Time how long the first one takes, including reading and understanding the instruction. Then time the fifth one. The fifth is faster because they’ve stopped re-reading the setup. Now give them a page where the instruction changes slightly every time, and time it again. It’s slower. That’s a cache miss, and they just measured it.

Ages 13+: Compute the break-even

Have your teen work the arithmetic from the real numbers. Base input $10 per million on Fable 5.1. Write = 1.25x, read = 0.025x. For a 20,000-token system prompt used N times in five minutes, compute total cost with and without caching, and find the N where caching wins. (It wins at N = 2.) Then have them redo it for the 1-hour cache at 2x write. Finish by reading the caching documentation’s section on cache invalidation and listing three things that would break the cache.

The question to ask: “What part of this stays the same every time, and what part changes? Which one should come first?”

Cached vs. not cached: the real cost table

Scenario	Tokens in the stable prefix	Cost per request (no caching)	Cost per request (with caching, Fable 5.1)	When caching wins
First request (cache write)	20,000	$0.20	$0.25 (1.25x write)	Never; you pay a premium
Second request within 5 min	20,000	$0.20	$0.005 (0.025x read)	Immediately
Ten requests in 5 min	20,000 each	$2.00	$0.295 total	By ~85% savings
Prompt under 512 tokens	400	$0.004	$0.004 (won’t cache)	Never; below minimum
Timestamp at the top of the prompt	20,000	$0.20	$0.25 every time	Never; cache never matches
Gap of 20 minutes between turns	20,000	$0.20	$0.25 (cache expired)	Only with 1-hour TTL at 2x write

Figures use Claude Fable 5.1’s published $10 per million base input price and the documented multipliers. Rates change; the structure doesn’t.

Look at the last two rows. Those are the failure modes, and both are entirely about where the changing content sits and how much time passes.

What to do at home

Keep the front of your prompt stable

If your kid uses a reusable prompt (a study template, a code reviewer, a writing coach), teach them to put the instructions at the top and the day’s specific question at the bottom. Every AI product that does this well is exploiting caching, and a kid who structures prompts this way is building a good habit regardless of price.

Finish a session in one sitting

The default cache lives five minutes. A focused 20-minute homework session with an AI tool is cheaper per question than the same 20 questions spread over an afternoon. That’s an argument for concentrated work sessions that has nothing to do with attention span and everything to do with arithmetic.

Don’t put the date at the top

This is the single most common mistake. “Today is November 4, 2026” at the start of a system prompt breaks the cache on every new day, and a timestamp with minutes breaks it on every request. Put changing details at the end.

Explain why the second answer felt faster

Cache hits reduce latency, not just cost. If your kid notices that follow-up questions come back faster than the first one in a session, that’s caching, and pointing it out turns a vague impression into an understood mechanism. It pairs naturally with test-time compute, which explains the opposite effect.

What not to do

Don’t tell your kid the model “remembers” things between conversations. It doesn’t. Caching stores a processed form of text for a few minutes on the provider’s servers to save computation. It’s not memory, it doesn’t persist, and it doesn’t mean the model knows anything about them later. Conflating caching with memory is how kids end up with wrong ideas about privacy.

What to Watch For Over the Next 3 Months

Week 4: Your kid can explain that only the unchanging front part of a prompt gets cached, and knows the default window is about five minutes.
Month 2 red flags: They think caching means the AI remembers them. Or they’ve started writing prompts with the question first and the instructions last, which defeats it.
Month 3 self-check: Show them two prompt structures and ask which one caches better. If they look for where the changing content sits, they’ve got it.

Frequently Asked Questions

Is prompt caching the same as the AI remembering our conversation?

No. Caching stores a processed form of prompt text on the provider’s servers for a short window, typically five minutes, purely to avoid recomputation. It isn’t memory, it doesn’t persist across sessions, and it doesn’t give the model knowledge about a user later.

How much did Fable 5.1 actually cut the price?

Cache reads dropped from $1.00 to $0.25 per million tokens, a 75% reduction announced September 1, 2026. Base input stayed at $10 per million and output at $50. The change comes from a lower multiplier: 0.025x base input instead of the standard 0.1x.

Why does the cache expire after five minutes?

Keeping processed prompt state in fast storage costs the provider money, so it’s held briefly by default. A one-hour option exists at 2x the write cost instead of 1.25x, which is worth it when requests are spread out rather than bunched.

Does caching help with the free version of an AI chatbot?

Indirectly. On consumer plans you don’t see token pricing, but the provider does, and cheaper repeated context is part of what makes generous free tiers possible. Caching is a supply-side economics change that shows up to families as better free products.

What breaks a cache?

Any change to the prefix before the breakpoint, including a timestamp or per-request data. Changing tool definitions, toggling web search or citations, or changing thinking parameters also invalidate it. And time: past the TTL, the entry is gone.

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

Sources

Anthropic. (2026). “Prompt caching.” Claude Platform Documentation. https://platform.claude.com/docs/en/build-with-claude/prompt-caching
Anthropic. (2026). “Pricing.” Claude Platform Documentation. https://platform.claude.com/docs/en/about-claude/pricing
MarkTechPost. (2026, September 1). “Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1: 52.6% on Terminal-Bench-Science and 75% cheaper cache reads.” https://www.marktechpost.com/2026/09/01/anthropic-releases-claude-fable-5-1-and-claude-mythos-5-1-52-6-on-terminal-bench-science-and-75-cheaper-cache-reads/
Anthropic. (2026, June 9). “Claude Fable 5 and Claude Mythos 5.” https://www.anthropic.com/news/claude-fable-5-mythos-5
Gizmodo. (2026, July 24). “Anthropic releases new Claude model, positions it as a cost-efficient version of Fable 5.” https://gizmodo.com/anthropic-releases-new-claude-model-positions-it-as-a-cost-efficient-version-of-fable-5-2000790486
Google. (2026, August). “Google AI updates, August 2026.” The Keyword. https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/
Snell, C., Lee, J., Xu, K., & Kumar, A. (2024). “Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters.” arXiv:2408.03314. https://arxiv.org/abs/2408.03314

Written by Ricky Flores

Founder of HiWave Makers and electrical engineer with 15+ years working on projects with Apple, Samsung, Texas Instruments, and other Fortune 500 companies. He writes about how kids learn to build, think, and create in a tech-driven world.