Anthropic released Claude Fable 5.1 on 1 September, alongside Mythos 5.1, a twin version with looser safeguards that is only available to vetted cybersecurity and life-sciences professionals. The headline for everyone else: Anthropic says Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads and up to approximately 45% less for highly agentic work.
Note what did not change. Input and output rates stay where they were. The entire saving comes from one line item: cache reads, the charge for re-reading context the model has already processed, drop 75% to $0.25/M tokens. If your firm has Claude embedded in delivery workflows such as report drafting, document review or spreadsheet agents, that pricing quirk matters more than the headline, because it decides whether you actually see the discount.
The other change worth a project director's attention is friction. Anthropic says its cybersecurity safeguards now intervene around 60% less often per session, and its biology safeguards fire 85% less often on benign medical and elementary biology queries. Fewer false positives means fewer agent runs dying halfway through a task for no good reason.
Why cache reads are the whole story
An agent working through a long task re-reads almost everything it has already seen on every turn, including the documents, instructions and its own earlier steps. Cognition, the company behind the Devin coding agent, measured a typical task at about 3 million cached tokens against just 21,000 output tokens. Over 95% of what an agent bills for is cache reads. Cut that rate by three quarters, and the task cost falls from roughly $5.00 to $2.68 on Cognition's benchmark, a 54% drop, without the model getting a penny cheaper on paper.
We're moving our Opus 5 traffic in Devin to Claude Fable 5.1 on launch day. With the new cache read pricing, a Fable-class model is finally economical for the workloads we'd kept on Opus, starting with code review.
The corollary cuts the other way: a workflow that sends short, one-shot prompts with little repeated context reads barely any cache and will see barely any saving.
Less friction on ordinary work
The safeguard changes are arguably the bigger practical shift for built-environment firms. Fable 5.1 can now be used to identify software vulnerabilities, though not to develop exploits, and the reduced false-positive rates should mean fewer document-review or data-analysis runs blocked because a classifier misread something benign. Anthropic's own prompting guide still lists situations that trigger false refusals, so the friction is reduced rather than removed.
Alongside this sits Enterprise Frontier Safeguards, a system Anthropic says gives enterprise customers zero-data-retention privacy by keeping data on infrastructure the customer controls, rolling out in phases from this autumn. For firms whose legal teams have blocked frontier models over data residency, that is the door opening, slowly.
Early testers liked the output quality too. "On our PowerPoint eval, Claude Fable 5.1 produced the best decks of any model we've tested, both in slide craft and in fully answering our research topic," said Aabhas Sharma, CTO of AI document platform Hebbia, in the same announcement.
The testimonials come from the vendor, of course, but they fit the audience for this release: firms producing reports, decks and reviewed documents at scale.
It's friendly Fable. Fable-level intelligence, Opus-level price, Sonnet-speed.
The caveat on your invoice
Independent testing complicates the cheerful maths. Artificial Analysis, which benchmarks models independently, scored Fable 5.1 top of its Intelligence Index at 66, but flagged it as very verbose: at maximum effort it generated 140 million output tokens across the index against a median of 71 million, and output still costs $50/M.
A model that writes twice as much can quietly eat a per-token discount, particularly at the higher effort settings. A per-token price cut is not the same thing as a cheaper completed task, and Cognition makes exactly that argument in its own analysis: measure cost per finished job, never cost per million tokens.
Anthropic's guidance points the same way. Its docs say Fable 5.1 at medium effort roughly matches Fable 5 at lower cost, and it defaults to high effort in Claude Code. If nobody in your firm has looked at effort settings since the model switched over, you are probably paying for thinking your tasks do not need it.
What to check this month
Pull the token breakdown on your Claude bill and check what share is cache reads, because that share is roughly the share of the discount you will actually receive.
Check which effort level your tools default to, and test whether medium or low holds quality on your routine drafting and review tasks.
Compare output token volumes per task before and after the switch, since a wordier model can offset the cache saving.
Recalculate your cost per completed task, not per million tokens, before renewing any budget assumptions built on Fable 5 numbers.
Log how often runs now end in refusals compared with August, so you can put a number on the reduced safety friction rather than taking it on trust.
Takeaway
This release rewards firms that run long, context-heavy agent workflows and punishes nobody, which makes it a rare pricing change worth acting on rather than just reading about. The 25% figure is Anthropic's estimate for typical workloads; yours will land somewhere between nearly nothing and 45% depending on how much context your agents re-read and how verbose the new model turns out to be on your tasks. The only way to know is to check the bill, and September is the month to do it.
We track what AI actually costs to run in a built-environment business, in money and in attention, every week at Project Flux. If your firm is trying to work out whether the model invoices match the vendor promises, subscribe free at projectflux.ai and keep the running costs honest.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
