Claude Fable 5.1 vs. Fable 5: On real work, I couldn’t tell them apart.
The New Stack Jessica Wachtel ● Covered by 5 sources
Anthropic’s Claude Fable 5.1 scored much higher on one benchmark. But in a set of hands-on tests, it looked a lot like Fable 5.
Based on reporting by The New Stack, Jessica Wachtel — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic launched Claude Fable 5.1 this week and pitched it as its most advanced model for coding and knowledge work. The splashy claim came with a big benchmark jump: on Terminal-Bench-Science, 5.1 scored 52.6% while Fable 5 scored 24.7%. Both models also cost the same, so that number was supposed to be the selling point.
I wanted to know whether that gap showed up outside the marketing slide. So I ran both models through four tasks that looked more like actual work: cleaning messy lab data, fixing a small Python project, solving two math problems, and auditing sensor readings with a pile of documented gotchas. All four tests were built from folders of data files with planted errors, which made the prompts useless to share on their own.
On the research task, both models got everything right in three turns. They excluded the same bad rows, handled a duplicated trial with a corrupted first entry, and produced the correct batch means. Fable 5.1 was a little faster, at 19.2 seconds versus 20.6, and slightly cheaper too, but the output was otherwise the same.
The coding test ended the same way. Each model found the obvious bug, where a remove function added stock instead of subtracting it, and the quieter off-by-one error in a threshold check. Both finished with 8 of 8 tests passing in three turns. Fable 5.1 got there faster, 13.6 seconds to 17.2, but again the practical result was a tie.
The reasoning problems were also a wash. Both models gave the correct answers and showed the right steps. 5.1 was a bit quicker and more concise, but not in a way that changed the experience. The hardest test was the sensor audit, where both models handled the fast clock, the hardware swap, the corrupted rows, and the Fahrenheit conversion correctly. Here, though, 5.1 took more turns, more time, and more money: 5 turns, 28.4 seconds, and $0.304, versus Fable 5’s 4 turns, 23.9 seconds, and $0.134.
Across all four tasks, both models hit 24 out of 24. The full run finished in 82.9 seconds for Fable 5.1 and 84.9 for Fable 5, but 5.1 used far more tokens and cost 34% more. So yes, the benchmark moved. The everyday difference did not really show up, and on the kind of work most people actually do, that matters more than a shiny percentage on a research test.
My take — AI-written commentary, not fact-checked reporting
This is the usual benchmark theater: one number goes up, everyone acts like the second coming has arrived, and real users are left squinting at the same outputs. If a model needs a lab-grade treasure hunt to look better, it’s not exactly screaming daily usefulness. The industry loves calling that progress; normal people call it a billing problem.
Read more about this at: The New Stack