Anthropic released Claude Fable 5.1 on September 1, 2026 saying it "will cost an estimated 25% less than Fable 5 for typical workloads," with savings up to roughly 45% on highly agentic work. In the week since, three independent measurements have been published, and they do not agree. One found the model costs about 20% more per task than its predecessor. Two found it substantially cheaper. All three are competently run, and none of them is wrong. The Fable 5.1 cost question has a single variable behind it, and it is one you can look up in your own billing in about ten minutes. This piece covers what Anthropic actually claimed and on what basis, the benchmark that found the model more expensive, the two that found the opposite, the number that reconciles them, why output grew, the API change that will break agent code regardless, and what to check before you forecast anything.
The short version: the discount is real and it is concentrated almost entirely in cache reads. If your workload is cache-heavy agentic work, the savings are larger than advertised. If it is not, the extra output tokens can make Fable 5.1 the most expensive option on the board.
The Fable 5.1 cost claim, and what it rests on
The pricing itself is unambiguous and we covered it when the model shipped: input stays at $10 per million tokens, output stays at $50 per million, and cache reads fall 75% to $0.25 per million. Nothing else moved. As we said in our launch coverage, this is not a price cut, it is a discount on one line item.
The 25% and 45% figures are estimates, and Anthropic has published very little about how it produced them. The only methodology detail we could find, reported secondhand rather than stated on Anthropic’s own pages, is that the numbers came from four weeks of Anthropic’s own August 2026 usage at default effort. We found no published figure for what share of that usage was cache reads, which is the single number that would make the estimate portable to anyone else.
That absence is the whole story. Hold onto it.
The benchmark that found it more expensive
Artificial Analysis runs a broad reasoning and knowledge suite it calls the Intelligence Index. Its published figures for maximum effort settings, which we pulled directly from its model pages:
| Model | Cost per Index task | Index score | Output tokens across the Index |
|---|---|---|---|
| Claude Fable 5.1 | $3.69 | 66 | 140M |
| Claude Fable 5 | $3.14 | 62 | 83M |
| Claude Opus 5 | $2.34 | 63 | not stated |
Fable 5.1 tops the index on capability and costs about 17% more per task than Fable 5 and about 58% more than Opus 5. It generates roughly 1.7 times the output tokens of Fable 5 to do it. Artificial Analysis summarized its own finding as the model costing "20% more per task than Fable 5 despite a 75% cache read price cut."
One disclosure belongs with those numbers. Artificial Analysis states that it "supported Anthropic with pre-release evaluation" of Fable 5.1, as it did for Fable 5. That makes it a pre-release partner rather than a fully arm’s-length party. It does not appear to have softened the result, which is unflattering to the vendor, and the figures are on its public pages where anyone can check them. But readers should know the relationship exists.
The two benchmarks that found the opposite
Both measured agentic coding rather than general reasoning, and both reached the reverse conclusion.
Cognition, on its FrontierCode benchmark, recorded Fable 5.1 at $2.68 per task at medium thinking against $5.84 for Fable 5, a 54% reduction, and cheaper than Opus 5 at $3.51. On verbosity it found the opposite of Artificial Analysis: Fable 5.1 "takes 33% fewer tokens to complete the same tasks relative to Opus 5," around 21,000 output tokens against 26,000.
Snorkel AI ran its own Terminal-Bench+ dataset and reported that Fable 5.1 "solved the tasks using 58% fewer output tokens and finishes 36% faster" than Opus 5. Snorkel flags this as a descriptive comparison rather than a fully controlled one.
So one benchmark has Fable 5.1 emitting 1.7 times the output of its predecessor, and another has it emitting a third fewer tokens than its stablemate. That looks like a contradiction. It is not.
The number that reconciles them
Cognition published the detail that explains everything: on its benchmark, "over 95% of the tokens on both models are cache reads."
Run the arithmetic. At a 95% cache-read share, almost every token in the workload just fell from $10 per million to $0.25 per million, a 97.5% cut on the dominant cost line. Output tokens are a small minority of the total and their price did not change, so even a meaningful increase in output volume is swamped. The model gets dramatically cheaper.
Now run it on a suite of discrete reasoning tasks with little repeated context. Cache reads are a small fraction of tokens, so the 75% cut applies to a small fraction of the bill. Output tokens are a large fraction, their price is unchanged at $50 per million, and there are now 1.7 times as many of them. The model gets more expensive.
Same model, same price sheet, opposite outcomes, and the variable is cache hit rate. Anthropic’s "typical workloads" phrasing quietly assumes a cache profile it never published, which is why the 25% figure is not portable.
This is the same trap we described when Fable 5’s metered pricing hit Pro subscribers, where output at $50 per million was what actually consumed the credit. The headline number described someone else’s usage pattern.
Why output went up at all
It is worth ruling out the obvious explanation: Anthropic did not turn the reasoning dial up. Its documentation gives the API default effort as high for both Fable 5 and Fable 5.1. No default changed between versions.
What Anthropic does document, on its own "what’s new" page, are two behaviors that cost output tokens. The model is "more likely to rewrite the entire file than make a targeted edit" when editing text files, and Anthropic notes plainly that "the rewrite costs more output tokens and time." It also describes "denser prose in places, with longer sentences and fewer paragraph breaks."
Set against that, the same page documents a change pushing the other way: "fewer progress updates during long tool runs," with less user-facing text between tool calls.
Which is honest of them, and also means Anthropic is not claiming net output fell. The higher token counts in Artificial Analysis’s numbers are a measured result, not something the vendor conceded.
The API change that breaks agent code
Independent of cost, one thing in this release requires code changes, and it is getting less attention than the pricing.
Forced tool use is gone. Anthropic’s migration guide states that on claude-fable-5-1, tool_choice values of {type: "any"} and {type: "tool", name: "..."} return a 400 invalid_request_error. If you built a router or a structured-extraction step that guarantees a specific tool gets called, that guarantee no longer exists.
Anthropic documents three replacements: leave tool_choice at auto, name the tool in the instruction and set strict: true; append a system message mid-conversation naming the required tool for that turn; or, if forced tool use was only ever a way to get schema-conformant JSON, use structured outputs instead. The first two are probabilistic where the old behavior was guaranteed, which is a real difference in an unattended loop.
There is a second change worth knowing. A Fable 5.1 thinking block "is valid only against the system prompt, tools, and conversation history that preceded it," so any application that edits or truncates history mid-conversation can trigger a 400. Anthropic enforces this for accounts created on or after August 31, 2026, and records but does not act on mismatches for older accounts, which means existing applications get a grace period rather than an immediate break. That is a good thing to have used before it expires.
What to actually check
Pull your cache-read fraction. In your Anthropic billing or usage export, find what share of your total tokens are cache reads. That one number tells you which of the benchmarks above describes you, and it takes minutes.
Above roughly 90%, expect savings at or beyond Anthropic’s estimate. Long-running agents with a large stable system prompt and tool set live here.
Below roughly 50%, model it yourself before assuming any saving. Discrete request and response work, classification, one-shot generation and most chat-shaped products live here, and the output-token growth may exceed the cache saving.
Either way, measure output tokens per completed task before and after, not per request. Cheaper per token and cheaper per finished job have come apart, and only the second one appears on an invoice.
And audit for tool_choice before you migrate anything. That is a code change with a deadline attached, not a configuration flag, and it is the part of this release most likely to page someone.
Frequently Asked Questions
Is Fable 5.1 cheaper than Fable 5?
It depends entirely on your cache hit rate. Input and output pricing are identical at $10 and $50 per million tokens; only cache reads fell, by 75% to $0.25 per million. On cache-heavy agentic work, Cognition measured a 54% cost reduction. On a general reasoning suite, Artificial Analysis measured Fable 5.1 costing about 20% more per task.
Why do the benchmarks disagree?
Because they measure different workload shapes. Cognition reported that over 95% of tokens on its coding benchmark were cache reads, so the 75% cache discount dominates. On a benchmark with little repeated context, cache reads are a small share of the bill, output tokens are a large share at unchanged pricing, and Fable 5.1 emits roughly 1.7 times as many of them.
What is the actual cost per task?
Artificial Analysis records $3.69 per Intelligence Index task at maximum effort, against $3.14 for Fable 5 and $2.34 for Opus 5. Cognition records $2.68 per FrontierCode task at medium thinking, against $5.84 for Fable 5 and $3.51 for Opus 5. Both are published figures on different workloads.
Did Anthropic publish how it got to 25%?
Barely. The only methodology detail we found, reported secondhand rather than on Anthropic’s own pages, is that the estimate came from four weeks of Anthropic’s own August 2026 usage at default effort. No cache-read share was published, which is the number that would make the estimate transferable to another workload.
Does Fable 5.1 think harder by default?
No. Anthropic’s documentation gives the API default effort as high for both Fable 5 and Fable 5.1. The extra output is attributable to documented behavior changes, notably that the model is more likely to rewrite a whole file than make a targeted edit, rather than to a change in reasoning settings.
Is Artificial Analysis independent?
Partly. It discloses that it supported Anthropic with pre-release evaluation of Fable 5.1, as it did for Fable 5, which makes it a pre-release partner rather than an arm’s-length third party. Its figures are published publicly and are unflattering to the vendor, and they can be checked directly on its model pages.
What broke in the API?
Forced tool use. On Fable 5.1, tool_choice set to type any or type tool returns a 400 invalid_request_error. Anthropic documents three replacements: auto plus a named tool in the instruction with strict set to true, a mid-conversation system message naming the required tool, or structured outputs where forced tools were only being used to get conformant JSON.
Does the tool_choice change break my code immediately?
The tool_choice rejection applies now. The related thinking-block binding check is enforced for accounts created on or after August 31, 2026, while older accounts have mismatches recorded but not acted on, which amounts to a grace period rather than an exemption.
Should I move from Opus 5?
Only after measuring your own cache profile. Artificial Analysis has Fable 5.1 at 58% more per task than Opus 5 on general reasoning, while Cognition has it cheaper than Opus 5 on agentic coding and Snorkel has it using 58% fewer output tokens than Opus 5 on terminal tasks. Nothing in that spread supports a decision made without your own numbers.
What is the one number to look up?
The share of your total tokens that are cache reads. Above roughly 90%, expect savings at or beyond Anthropic’s estimate. Below roughly 50%, model the change yourself, because the output-token growth may exceed the cache saving.
Compliance note. Pricing, the effort default, the documented behavior changes, the tool_choice rejection and the thinking-block binding rules are quoted from Anthropic’s own model documentation and migration guide. The 25% and 45% figures are Anthropic estimates and the methodology behind them is reported secondhand, not published by Anthropic. Artificial Analysis figures were pulled from its public model pages and it discloses a pre-release evaluation relationship with Anthropic. Cognition and Snorkel AI figures are each vendor-published results on their own benchmarks, and Snorkel describes its comparison as descriptive rather than fully controlled. No independent verification of Anthropic’s "typical workload" framing has been published by any party.
