Fable 5.1 more than doubles Fable 5 on Terminal-Bench-Science, 52.6 percent against 24.7
Cache reads drop from 1.00 to 0.25 dollars per million tokens, a 75 percent cut that input and output pricing do not show
It beats Opus 5 on all nine published benchmarks, but the gaps range from 0.4 points to 23.6
The cheapest way to read this: pick the benchmark closest to your actual work, ignore the rest
Anthropic shipped Claude Fable 5.1 on 1 September 2026. Nine benchmarks came with it, and the headline number is genuinely large: Terminal-Bench-Science 0.1 goes from 24.7 percent to 52.6 percent.
The rest of the table is more interesting than the headline, because the gaps are wildly uneven. Here is what the numbers actually say.
The full table, in one place
Every figure below is from Anthropic's own launch page. Fable 5.1 against its predecessor, against Opus 5, and against GPT-5.6 Sol where a competitor number was published.
Benchmark
Fable 5.1
Fable 5
Opus 5
GPT-5.6 Sol
Terminal-Bench-Science 0.1
52.6%
24.7%
29.0%
22.4%
Terminal-Bench 4.0
55.8%
42.0%
52.3%
37.3%
GDPval-AA v2
1853
1723
1824
1711
OSWorld 2.0 (partial)
77.9%
72.9%
75.4%
not published
OSWorld 2.0 (strict)
41.7%
36.1%
39.6%
not published
Humanity's Last Exam (no tools)
60.9%
57.8%
56.6%
not published
Humanity's Last Exam (with tools)
65.0%
63.8%
63.6%
not published
AutomationBench
31.4%
17.1%
26.9%
19.6%
CursorBench 3.2.0
73.4%
70.5%
70.0%
67.2%
Fable 5.1 wins every row. That is the easy reading and it is also the useless one.
The gaps are not the same size, and that is the whole story
Sort those rows by how much Fable 5.1 beats Opus 5 by and two clusters fall out.
The blowouts are Terminal-Bench-Science, 23.6 points, and AutomationBench, 4.5 points on a base of 26.9, which is a relative jump of roughly a sixth. Both measure long agentic runs where the model has to keep going without a human correcting it.
The rounding errors are Humanity's Last Exam with tools, 1.4 points, and OSWorld strict, 2.1 points. GDPval-AA moves 29 points on a scale where Fable 5 already sat at 1723, which is under two percent.
So the shape is consistent. If your work is a single hard question, Fable 5.1 is barely distinguishable from Opus 5 at a fifth of the input price. If your work is an agent grinding through forty tool calls unattended, the difference is not subtle.
The Terminal-Bench-Science number deserves one caveat. Doubling from 24.7 to 52.6 is enormous, and it is also a benchmark where everyone was scoring badly. Going from "fails three quarters of the time" to "fails half the time" is real progress and still not a number you would build an unsupervised pipeline on.
The pricing change is the part that is easy to miss
Input stays at 10 dollars per million tokens. Output stays at 50. Anthropic lists in dollars, so those are dollars, not euro. Read only those two lines and you would conclude nothing changed.
Cache reads went from 1.00 to 0.25 per million tokens. That is a 75 percent cut, and Anthropic puts the practical effect at roughly 25 percent off typical workloads and up to about 45 percent off highly agentic ones.
That number is only real if your cache actually hits. Prompt caching is a prefix match, so any byte that changes early in the prompt invalidates everything after it. A timestamp in your system prompt, an unsorted JSON blob, a tool list that reorders itself between requests: each one silently drops your hit rate to zero, and the bill looks identical to someone who never enabled caching at all.
The check is one field. Log usage.cache_read_input_tokens across repeated requests. If it stays at zero, you are paying full input price and the 45 percent headline is not yours. I wrote about the same class of silent failure in The Error Log I Read Every Morning, where the expensive bugs were the ones that never raised.
The science numbers are the strangest part of the launch
Alongside the benchmarks, Anthropic published three results from real scientific work, and they are a different kind of claim entirely. A benchmark is a score. These are outcomes.
In protein design, models produced binders with roughly ten times higher binding affinities than the previous competition results, with a hit rate near 50 percent across twelve targets. For context, a hit rate of a few percent is normal in that field, so half is not an incremental improvement.
In planetary science, radar data of Venus was reprocessed to reveal detail down to two to three kilometres rather than the previous ten to twenty. That is the same raw data, read better.
In GPU work, the model found speedups of up to 2.5 times across seven deep learning models, with an estimated 30 to 60 percent cost saving on genome-wide analyses.
Treat all three the way you would treat any vendor case study. They are unreplicated, they are selected, and none of them come with the failed attempts alongside. What makes them worth mentioning anyway is that they are falsifiable in a way a benchmark percentage is not: a binding affinity either reproduces in someone else's lab or it does not.
The practical read is narrow. If your work looks like these, long autonomous analysis over a large dataset where the answer is checkable at the end, the agentic benchmark gaps and these case studies are pointing at the same thing. If your work is a chat feature, they tell you nothing at all.
Where 5.1 does not help you
Anthropic published its own limitations, which is worth reading before you migrate anything.
Their automated behavioural audit has "less visibility into very long-context work and multi-agent settings", and less coverage of impossible tasks than they would like. Both of those are exactly the shape of the workloads the agentic benchmarks say 5.1 is best at. The model is strongest where the safety testing is thinnest. That is not a scandal, it is a known gap stated openly, but it means a long autonomous run still needs your own evals rather than a benchmark score.
There is also a change that has nothing to do with capability. New API accounts can no longer manually edit Claude's prior context in a multi-turn conversation while preserving the transcript. That is a distillation countermeasure, and if you built tooling that rewrites assistant turns to steer a conversation, it will stop working on new accounts.
Mythos 5.1, the same model with different safeguards, scored 60.9 percent on Terminal-Bench 4.0 against Fable 5.1's 55.8. You cannot have it. It is gated behind the Cyber and Life Sciences verification programmes and currently limited to a set of US organisations.
What I would actually change in a running project
Three things, in the order I would do them.
First, turn on caching properly before touching the model ID. The 45 percent figure is a cache figure, not a model figure. If your hit rate is already high, switching to 5.1 hands you most of that saving on day one. If it is zero, you get nothing and you will blame the wrong thing. Fix the prefix, confirm cache_read_input_tokens is non-zero, then migrate.
Second, re-tune effort rather than assuming your old settings carry over. Effort matters more on this generation than it did before, and lower effort on a newer model frequently beats high effort on an older one. That is the cheapest experiment available: same code, one string, measured on your own traffic instead of someone's leaderboard.
Third, handle the refusal path before it hits production. Fable returns HTTP 200 with stop_reason: "refusal" and a category in stop_details. Code that reaches straight for content gets an empty string and shows a blank response to a user. The server-side fallback parameter routes those requests to another model automatically, which is a two-line change and considerably better than discovering the failure from a support email.
None of that requires a rewrite. It is a config pass, and most of the saving lives in it rather than in the model swap everyone leads with.
Bottom Line
Pick the row closest to your actual work and ignore the other eight. That is the honest way to use a launch table.
If you are doing single-shot reasoning, Opus 5 remains the value pick and the Fable premium buys you very little. If you are running agents that work unattended for a long time, Terminal-Bench-Science and AutomationBench say the gap is real and large, and the cache read cut makes the price difference much smaller than the sticker suggests.
The migration itself is small. The model ID is claude-fable-5-1. Thinking is always on, so omit the parameter rather than configuring it. Control depth with output_config.effort instead. Set up a refusal fallback while you are in there, because Fable can return stop_reason: "refusal" with a 200, and code that reads content without checking will hand your users an empty string.
If you want the setup that keeps this kind of thing straight across projects, Claude Blueprint is the configuration I ship with. And the rest of the model coverage lives in the Lab Overview.
Top comments (0)