In the same week that Claude Fable 5.1 shipped, two things happened that look like they cannot both be true.
Millennium had a piece of code that crashed about one run in a million. Their team had not explained it in four to five years. Every model they tried missed it, Fable 5 included. Fable 5.1 found it.
That same week a creator titled his review "I'm asking for a refund", and Artificial Analysis, which ranks these models for a living, put Fable 5.1 at the top of its intelligence index while measuring it at 20% more cost per task than Fable 5.
Both are accurate, and the reason is not in the benchmarks. It is in the pricing page and in how you pay for the model.
What I actually did
I did not run my own benchmark suite. There are plenty of those already and they mostly disagree.
What I did was read the release the way you would read a spec sheet before buying hardware: the announcement, the model docs, the system card, the prompting guide, the pricing page, and the support articles about plan limits. Then the independent measurements from Artificial Analysis, ARC Prize, FrontierSWE and Simon Willison, because the vendor's own table is not evidence about the vendor.
That turns out to be the differentiator. Almost every review this week ran demos and picked a side. The interesting material was sitting in the footnotes.
The cut is one line item
Fable 5.1 is not broadly cheaper. Input is still $10 per million tokens and output is still $50, unchanged from Fable 5.
One line moved. Cache reads, meaning the tokens the model rereads from a prefix it already processed, went from $1 to $0.25 per million. That is the entire price cut.
It matters more than it sounds, because in a long agent session most of your tokens are re-reads of the same context. Anthropic measured four weeks of real August traffic and reports roughly 25% off a typical bill and up to 45% off a heavily agentic one.
It also produces the one situation where the expensive model undercuts the cheaper one:
Cache read, per million tokens
Fable 5.1 $0.25
Opus 5 $0.50
Everything else on Fable costs double what Opus costs. A cache heavy loop really can be cheaper on Fable. Nothing else can.
Why the per-task number still went up
The second dial is effort, and it moves the bill much harder than the cache line does.
Simon Willison drew the same pelican at every effort level. Low cost him 10 cents. High cost 13 cents. Xhigh cost $1.83. Max cost $3.30, producing 65,927 output tokens over 14 minutes. Same prompt, 33 times the price.
Artificial Analysis runs everything at max. So they measured $3.76 per task against $3.14 for Fable 5, which is the 20% increase people are quoting. The cache cut is why it was not worse: without it the same measurement would have landed at $5.16.
So the honest version of the headline is that the model got cheaper to reread and more expensive to think, and which of those dominates depends on your workload. If you are running one shot requests at max effort, you are the case they measured. If you are running a review loop over the same repository for two hours, you are the case Anthropic measured.
The plan is a different meter entirely
None of that per-token math applies if you are on a subscription, and this is where the refund videos come from.
On Pro, Fable is not part of the plan's usage limits at all. It runs on pay as you go usage credits, and unlike previous launches there is no one-time credit this time.
On Max it is included, for up to 50% of your weekly limits. The same support page adds that these models consume those limits faster than other Claude models. That sentence is doing a lot of work and almost nobody quotes it.
Then the counter itself. The pricing page says Max gives 5x or 20x more usage per 5-hour session than Pro. Per session, with weekly limits stacked on top. A lawsuit filed June 15 alleges the real weekly multipliers are far below the advertised ones. And on September 14 a permanent 25% raise of weekly limits takes effect, which Anthropic itself clarified works out to a 17% reduction compared to today, because a temporary 50% boost ends the day before.
The model is not more expensive per token on a plan. It spends a budget faster than that budget was ever advertised to be spent.
Three things break if you call the API today
If you have a working Fable 5 integration, three changes fail it rather than degrade it.
Forced tool use returns a 400. Setting tool_choice to any or to a named tool is rejected. Thinking is always on now, and forcing a call would skip it.
Thinking blocks record which model wrote them, one way. Fable 5.1 reads Opus's thinking. Nothing reads Fable 5.1's.
Conversations are append only. Edit an earlier turn, or rebuild your system prompt or tools array between requests, and the next call fails with a message about the block being bound to a different conversation. This is enforced for accounts created on or after August 31. It breaks every harness that injects a reminder and deletes it on the following turn, which is a very common pattern.
Five things arrive at the same time, and the useful one is that you can change effort per message without losing your cache.
Then there are the documented behavior regressions, which are the ones burning people's quota right now. It rewrites whole files for small edits. It issues one tool call per turn in coding loops where Fable 5 batched several. It sometimes describes what it would do next instead of doing it. Each of those has a one line fix in the prompting guide, and each of them costs output tokens until you apply it.
What this doesn't prove
This is a documentation read, not a capability test. I cannot tell you whether the model is better at your work, and I did not try to.
The benchmark table is also weaker evidence than it looks. Anthropic's own footnote gives a standard error of 3.5 to 4.5 points per model, which puts the three point coding lead over Opus 5 inside the error bar. The rows missing from the marketing graphic are the ones where Opus 5 wins: 90.42 against 90.0 on ARC-AGI-2, and 59.4 against 54.7 on SWE-bench multimodal.
One more thing worth knowing before you send anything sensitive: Fable 5.1 carries 30-day retention and is not available under zero data retention unless Anthropic authorizes it. Enterprise Frontier Safeguards is the way out, and it ships this fall.
The full run
The video walks the actual pages, so you can see where each of these numbers lives rather than taking my word for the quote.
If you are on Max: did your 5-hour window survive the first week of Fable 5.1, and did you change your prompts before or after finding out?
English is not my first language, so I use a tool to polish the writing.
Top comments (1)
This framing — "read the release like a spec sheet before buying hardware" — is exactly right, and the cache-read line is the one most people skim past. The thing I'd add from running long agent sessions: the 25–45% savings only shows up if your prefix is actually stable. The moment your orchestrator reorders tools, injects a timestamp, or rotates a system-prompt fragment, you bust the cache and quietly pay full input price on the re-read. We've watched "identical" agent loops differ 3–4x in bill purely on prompt determinism.
The effort dial point is the sharper one, though. Benchmarking everyone at max is defensible for a leaderboard but it's the worst possible proxy for production, where most tasks want low/medium and you reserve xhigh for the one gnarly step. Curious whether you think the per-task number should be reported as a curve across effort levels rather than a single figure — a scalar there feels almost designed to mislead.