Anthropic shipped Claude Fable 5.1 on 1 September. The headlines are about benchmarks and who it beats. If you have an AI-backed process actually running inside your company — or you're thinking about building one — the figure that will change your bill isn't that one: it's that cache reads dropped by 75%.
It sounds like plumbing detail. It isn't, and here's why.
Why the cache is what drives the bill
A language model has no memory between calls. Every time you talk to it, you send the whole conversation again: the instructions, the documents, everything already said. All of it, from the top.
Which means that in an automated process the cost isn't driven by how long the answer is, but by how many times the cycle repeats. A process that reads an invoice, classifies it, checks whether the result adds up and corrects it when it doesn't, makes several calls in a row. Every one of them re-sends the same context.
The cache exists for exactly that: the part that doesn't change is stored, and re-reading it is charged at a much lower rate. When the provider cuts the price of that re-read by 75%, it isn't touching a small slice of your spend. It's touching the biggest one.
Anthropic's numbers: around 25% less on typical workloads, and up to 45% less on highly agentic work — that is, processes that chain many steps on their own. Base input and output pricing doesn't move; the cache does.
What this changes in practice
Things that didn't add up before start to add up. A process that checks each document twice instead of once, or that verifies its own work before handing you a result, costs twice the calls. With cache reads at a quarter of the price, "it costs twice as much" stops being the argument that kills the idea in the meeting.
It's the difference between an AI that answers and an AI that works: the second one needs several passes, and until now every pass was expensive.
The small print almost nobody is reading
This is where the press release stops being useful and you have to go to the documentation. Three things that matter more than any benchmark if you're putting this inside a company process:
1. Some queries get routed to a weaker model on their own. Anthropic says so explicitly: queries flagged by its safeguards in cybersecurity and biology are automatically routed to less capable models. If your process brushes against those areas — and an IT support process brushes against them more often than you'd think — quality can change without you touching anything and without anyone telling you. You check that before you build, not after.
2. Data is retained for 30 days by default, for safety monitoring. That's neither good nor bad: it's a decision you're making, one you need to document and one you need to tell your client about if their data goes through it.
3. A benchmark doesn't tell you whether it works on your thing. A model gaining points on a public test doesn't mean it reads your invoices better, in your format, with your awkward edge cases. You only find that out by running it against your own documents.
It's also been reported that Claude Code users should see around 60% fewer cybersecurity false positives. That's a real and welcome improvement, but note what it is: a figure for one specific product, not a general promise about your process.
What this does not fix
Cheaper tokens don't turn a bad process into a good one. If nobody today knows where the invoices are, who approves them, or on what criteria, a cheaper model will just hand you the same mess faster and for less money.
The order is the same as always: understand the process first, automate the part that can be automated second, and the model is the tool for the last step, not the starting point. What this price cut changes is that the cost conversation stops being the brake — and that's already quite a lot.
Any questions, tell me and we'll look at it.