claude fable 5.1 is cheaper. the discount is all in one line item.

elisabeth hitz · published september 3, 2026 · updated september 3, 2026 · 7 min read

anthropic shipped a price cut two days ago. it is real, and it is narrower than the headline suggests. it lands on exactly one line item, and most people are set up to miss it.

claude fable 5.1 did not get a lower base price. anthropic cut the cache read rate from 0.1x to 0.025x of base input, a 75% drop, on september 1, 2026. that works out to roughly 25% off a typical workload and up to 45% off heavily agentic work. you only collect it if you reuse the same context.

that distinction is the whole post. a discount on cached tokens is not a discount on using claude. it is a discount on not retyping yourself. if your setup is a fresh chat and a fresh paste every time, your bill did not move.

what anthropic actually changed on september 1

anthropic released claude fable 5.1 and claude mythos 5.1 on september 1, 2026, and the pricing change was entirely on cached input. the two models are the same underlying model with different levels of safeguards (Anthropic, sep 1 2026). the headline benchmark: fable 5.1 scores 55.8% on terminal-bench 4.0 against 42.0% for fable 5 (Anthropic, sep 1 2026). claude code users should also see around 60% fewer cybersecurity false positives (Anthropic, via MacRumors, sep 1 2026).

but the number that changes what anything costs you is this one. on most claude models a cached input token bills at 0.1x the base input rate. on fable 5.1 and mythos 5.1 it bills at 0.025x (Anthropic developer docs, sep 2026). anthropic puts the practical effect at roughly 25% less than fable 5 for typical workloads, and up to 45% for highly agentic work (Anthropic, sep 1 2026).

the cheaper line item is your context, not the model

one row on the price sheet moved. everything else is exactly where it was. here is the full picture, as multiples of the base input rate.

read that table as a sentence and it says something useful: anthropic made it cheaper to re-read what you already gave it, and left the price of telling it something new alone. that is a pricing signal about how they think the work should be structured, and it points the same direction i have been pointing for a year. the standing context belongs in a file, not a chat.

what a cache read is, in plain terms

the model does not remember anything between requests, so every turn re-sends the whole conversation from the top. prompt caching is how the api avoids reprocessing the part that did not change. it matches the start of each request, the prefix, against content it recently processed. the match is exact, so a change anywhere in the prefix recomputes everything after it (Anthropic developer docs, sep 2026).

two details decide whether this ever pays you.

• the 512 token floor. on fable 5.1 and mythos 5.1 the minimum cacheable prefix is 512 tokens. below that, nothing is cached, and the api returns no error to tell you so (Anthropic developer docs, sep 2026). a short throwaway prompt collects none of this.
• the cache expires on idle. the default time to live is five minutes, with a one hour option that bills cache writes at 2x base input instead of 1.25x (Anthropic developer docs, sep 2026). walk away long enough and the next turn pays full price to rebuild.

so the discount rewards a specific shape of work: a substantial context you keep stable, reused in a session you keep warm. that is not how most people use claude. it is how an operator uses claude, which is the difference i wrote up in context engineering for operators.

on the accounts i have set claude up on, the ones that got cheaper were never the ones with better prompts. they were the ones that stopped retyping their context and put it in a file the model reads the same way every session.

five things that throw the discount away mid-session

if you work in claude code, caching is already on and needs no configuration. the way you lose the discount is by changing something early in the request, which forces a full reprocess at base price. these are the five that come up most, all from anthropic's own list.

1. switching models. each model has its own cache. switching with /model re-reads the entire conversation with no cache hits, even when the content is identical.
2. changing effort level. the cache is keyed by effort as well as model, so /effort mid-session does the same thing.
3. turning on fast mode. it adds a request header that is part of the cache key, so the first turn after you enable it reads everything uncached. this costs less at the start of a session than deep into a long one.
4. connecting or disconnecting an mcp server whose tools load into the prefix rather than being deferred. tool definitions sit in the system prompt layer, so changing the set invalidates everything below it.
5. running /compact. it replaces your history with a summary, which by design does not share a prefix with the old one.

the practical rule anthropic gives is one line: pick your model and effort at the top of a session, and save /compact for natural breaks between tasks. worth knowing what is safe, too. editing files in your repo, invoking skills and commands, changing permission mode, and /rewind all keep the cache intact.

one trap that catches people: editing CLAUDE.md mid-session does not break the cache, but the edit also does not apply. claude keeps working from the version loaded at session start, and the new content loads on the next /clear, /compact, or restart (Anthropic Claude Code docs, sep 2026). if you have been editing your context file and wondering why nothing changed, that is why. the file itself is worth getting right first, which is what claude.md for founders is for.

how to earn the discount, in five steps

the setup that collects this is the same setup that makes claude useful, which is convenient. in order:

1. move the standing context into a file. the things you retype every session (who you are, what the business does, how you want output shaped) go into CLAUDE.md, where they sit in the stable project layer instead of the conversation.
2. clear 512 tokens. that is the floor for caching on fable 5.1. a thin file caches nothing. this is the one case where more context genuinely helps, which is not the usual case, as i argued in more context is not better.
3. choose your model and effort before you start, not three turns in.
4. batch the cache breaking moves. mcp changes, fast mode, and /compact at a break between tasks, never mid-task.
5. check the hit ratio. run /usage and read the prompt cache line in the session block. a high read to creation ratio means it is landing. creation staying high every turn means something in your prefix keeps changing.

fable 5.1 vs mythos 5.1: the one you cannot choose

they are the same model. fable 5.1 is generally available; mythos 5.1 has more permissive safeguards for cybersecurity and life sciences work and ships only through anthropic's trusted access programs, limited to vetted us companies and individuals (Anthropic, sep 1 2026). so if you are a solo builder or a small business, this is not a decision you get to make. you are on fable 5.1, and the pricing above is your pricing.

that split is worth noticing but not worth agonising over. capability tiered by who you are rather than what you pay is now a normal shape in this market, and i covered what it means for people outside the gate in what builders must do when access is controlled and the gated sol model. the short version has not changed: build on what is generally available, and do not architect around access you do not have.

the takeaway

the honest read of september 1 is that anthropic did not make claude cheaper. it made a particular way of working cheaper by 75%, and left everything else at list price. the way it rewards is the one where your context lives in a file, gets reused, and stays stable across a session.

if you already work that way, you got a raise this week and you had to do nothing. if you work in fresh chats and fresh pastes, nothing happened to your bill, and the fix is not a better prompt. it is a context worth caching. that is a one afternoon job, and it pays twice: once in output quality, once now in price.

common questions

why is claude fable 5.1 cheaper than fable 5 if the input price is the same?

because the cut is on cache reads, not on base tokens. on most claude models a cached input token bills at 0.1x the base input rate. on fable 5.1 and mythos 5.1 it bills at 0.025x, a 75% reduction. base input, base output, and both cache write rates are unchanged. anthropic puts the practical effect at roughly 25% off a typical workload and up to about 45% off highly agentic work, because agentic work re-reads a large stable prefix on every turn.

how do i actually get the cache read discount in claude code?

you already get it, as long as you do not keep breaking the cache. claude code caches the system prompt, tool definitions and project context automatically, with no configuration. the discount is lost when something early in the request changes: switching models, changing effort level, turning on fast mode mid-session, connecting or disconnecting an mcp server whose tools load into the prefix, or running /compact. pick your model and effort at the start of a session and save /compact for natural breaks.

what is the minimum prompt length for prompt caching on claude fable 5.1?

512 tokens. on fable 5.1, mythos 5.1, opus 5, fable 5 and mythos 5 the minimum cacheable prefix is 512 tokens. anything shorter is simply not cached, and the api returns no error to tell you so. this is why short throwaway prompts see none of the discount: there is nothing large enough to cache.

can i use claude mythos 5.1, or only fable 5.1?

almost certainly only fable 5.1. anthropic says the two are the same underlying model with different levels of safeguards. fable 5.1 is generally available. mythos 5.1 has more permissive safeguards for cybersecurity and life sciences work and is released only through trusted access programs, limited to vetted us companies and individuals. for a solo builder or a small business, mythos is not a choice on the menu.

building this out with other people is easier than doing it alone. the ai builders lounge is where the weekly builds happen.

or just follow along. new field notes most weeks on x, instagram, and tiktok.

browse all 30 free tools →