Key Takeaways

• Anthropic cut cache-read pricing 75 percent on September 1, 2026 — good news for efficient usage, easy to overlook otherwise.
• A cache miss on the newest models now costs 40 times what a cache hit costs, versus 10 times on every other Claude model.
• Claude subscriptions quietly shrink your cache lifespan from one hour to five minutes once you're spending paid usage credits.
• Switching models, switching effort level, resuming stale sessions, and mid-task compaction all wipe your cache — and are all avoidable.
• Run /cost and read the prompt cache line. It now tells you the hit ratio, the miss count, and — as of version 2.1.260 — why the misses happened.

On September 1, 2026, Anthropic cut the price of a specific kind of Claude usage by 75 percent. If your team uses Claude or Claude Code for anything — writing, research, coding, automations — that sounds like straightforward good news.

It isn't quite that simple. The same change that made efficient usage nearly free also made one particular mistake four times more expensive than it used to be. Here's what actually changed, translated out of engineer-speak, and the one 90-second check that tells you which side of this you're on.

What did Anthropic actually change on September 1?

Anthropic cut the price of "cache reads" — a cheaper way of reprocessing context Claude has already seen — from about a dollar per million tokens down to $0.25 per million tokens, as part of the launch of Claude Fable 5.1 and Mythos 5.1 (Anthropic, Sept 1, 2026). For most people, that shows up as roughly 25 percent off the total bill. For heavier, repetitive automation work, it can run closer to 45 percent off (source).

The technical detail that matters for your wallet: on the new model, a cache read now costs 0.025x the price of regular input, compared to 0.1x on every other Claude model (Claude Platform release notes). That single ratio is the whole story of this post.

What is a "cache," in plain English?

A cache is Claude's way of not re-reading everything from scratch every time you send it a message.

Picture a long conversation, or a coding project with a big set of instructions and files attached. Every time you send a new message, Claude technically has to process all of that background context again — unless it's cached. A cache is like a sticky note Claude keeps nearby with all that context already digested. If the note is still there, reusing it (a "cache hit") is dramatically cheaper than reprocessing everything from zero (a "cache miss"). If the note has been thrown away, you pay full price to write a new one.

What's a "token"?

A token is the unit AI companies bill by — roughly three-quarters of a word. A million tokens ("MTok") is a lot of text, so prices are usually quoted per million.

What's a "TTL"?

TTL stands for "time to live." It's how long that sticky note stays valid before Claude throws it out and has to write a fresh one — at full price — the next time you ask something.

Why did a price cut make mistakes more expensive?

Because the price cut only applied to cache reads, and left regular, uncached input right where it was. Regular input still runs about $10 per million tokens. A cache read on the newest model is now $0.25 per million tokens (Claude Platform release notes).

Do the math and you get a 40-to-1 penalty for a cache miss on the newest model, versus 10-to-1 on every other Claude model.

In other words: Anthropic made cached work nearly free and left uncached work at full price. That means your bill is now dominated less by how much you use Claude, and more by how often the cache gets broken along the way.

What's the hidden trap in a Claude subscription?

The trap is that your cache doesn't always get the same amount of time to live. On a Claude subscription, the main conversation gets a one-hour cache TTL by default. But the moment your usage exceeds what's included in your plan and starts drawing on paid usage credits, that TTL quietly drops to five minutes — twelve times faster expiry, arriving at exactly the moment you start paying cash per token.

This is documented, expected behavior — the longer, one-hour cache option bills its initial "write" at a higher rate, and Anthropic doesn't spell out the reasoning further (How Claude Code Uses Prompt Caching). But almost nobody realizes it's happening, which means almost nobody is checking for it.

How do I check my own cache performance in 90 seconds?

Run the /cost command inside Claude Code and read the "Prompt cache (main)" line. As of version 2.1.251 (shipped August 28, 2026), that line reports your hit ratio, your miss count, how many tokens got re-cached, and whether your cache is warm or cold. As of version 2.1.260 (September 3), it also tells you the likely cause of any misses (Claude Code changelog).

That's it. One command, under two minutes, and you know exactly where your money is going instead of guessing.

We run this check on our own Claude Code sessions at vixxiv every week — not because it's fun, but because it's the difference between a bill that makes sense and one that doesn't. Twenty-seven years of building eCommerce operations from scratch teaches you to check the meter; it's the same instinct that has us still elbow-deep in the newest AI tooling today, building our own automations instead of reading about someone else's.

[NEEDS FROM JOE: screenshot of an actual /cost output showing the prompt cache hit ratio, miss count, and likely-cause line]

What four habits silently blow up your cache mid-task?

Four specific, everyday actions reset your cache and put you back at full price, according to Anthropic's own documentation (source):

• Switching models mid-session. Every model keeps its own separate cache. Swap models and you start over at full price.
• Switching effort level mid-session. Same rule applies — every effort setting has its own cache. Pick your model and effort level at the start of a session and leave them alone.
• Resuming a long session after an update. A background update changes the underlying system prompt, which means the entire history has to reprocess uncached.
• Using a mid-task summary-and-compress command. It rewrites your history, which breaks the cached prefix. Truncating back to an already-cached point, or appending a recap instead of replacing history, keeps the cache intact.

"The first turn back into a long session can be the most expensive request you send." That's Anthropic's own documentation talking about resuming stale sessions — and for most teams, that first turn happens on Monday morning.

Is there a risk hiding in your company's AI gateway?

Yes, if you route Claude usage through a corporate LLM gateway rather than straight to Anthropic. Some gateways strip out the technical markers that tell Claude what to cache, while still returning a normal success response — no error, no warning. If that's happening, your entire conversation bills as fully uncached input on every single turn.

That's a silent 40-to-1 tax on every request, and the first sign of it is usually just a much bigger invoice.

What should I actually do about this?

Start with the 90-second check, then fix what it shows you.

• Run /cost and read the prompt cache line. That's your baseline.
• If you're on an API key, a cloud provider, or you routinely exceed your plan into paid usage credits, set promptCacheTtl to the one-hour option. Subagents run on a separate, always-five-minute cache bucket, so set subagentPromptCacheTtl too if you use them. Both settings require Claude Code version 2.1.242 or later.
• Agree on a model and effort level before a work session starts, and stick with it.
• Treat Monday mornings — or any first message back into a long-idle session — as the priciest moment of your week, and budget for it rather than being surprised by it.
• If your team uses a corporate gateway, ask whoever manages it to confirm cache markers are passing through intact.

None of this requires anyone on your team to become an engineer. It requires someone to run one command and read one line.