Terminal-Bench-Science hands a model a shell and a set of multi-step research tasks, then scores the fraction it completes correctly [3]. A pass rate is fundamentally a property of the task distribution being sampled. The 27.9 points between the two scores [17] therefore sit entirely in tasks Fable 5 fails. On any task Fable 5 already finishes, the headroom is zero and 5.1 can only match it.
Had the four fixtures been drawn from the benchmark's own distribution, a model that completes 24.7% of it [2] would sweep all four about 0.4% of the time [18]. The fixtures instead sit outside that distribution, and that gap is the finding. The planted bugs sat behind a suite that failed loudly, and the bad experiment rows were described in lab notes the model was free to read [10][8]. Both fixtures are well specified. The Terminal-Bench-Science tasks Fable 5 misses are, by construction, the ones it does not finish.
Two conditions have to hold before that benchmark delta reaches a pipeline. The queue needs a real population of work Fable 5 currently fails, and each task needs enough repeats to show a rate rather than an outcome. One run per task cannot separate 24.7% from 52.6%, because both numbers predict passes and failures on the same task.
The measured differences are small and consistent. On the coding fixture 5.1 finished 3.6 seconds sooner, 21% off a 17.2-second run, at the same seven cents [11][19]. It saved 1.4 seconds and 1.4 cents on the research fixture, which comes to $14 per thousand runs [9][20]. The two reasoning problems cost it 425 fewer output tokens, 23% fewer, worth about two cents at $50 per million [12][21]. With pricing identical [4], none of that argues against switching. It argues against rebuilding anything to mark the occasion.
Dan Shipper of Every called 5.1 the strongest coding model his team has used, after a week of testing [5]. That holds alongside a four-task tie without much strain: a week of varied production work samples far more of the hard tail than four fixtures and a deliberately harder tiebreaker do [13], and the tail is where the 2.1x lives [16]. The New Stack's own caution stops short of an accusation. It notes that benchmark sets are scored under conditions vendors help define, and that tuning toward graded tests has happened, while declining to claim Anthropic did it here [15].
The weak point of the comparison is reproducibility: the prompts are useless without the folders of planted-error data behind them, so they were not published [14]. Take it as one workload, run once, across four tasks. That is still a better predictor of your own queue than a vendor table, and it is the shape of test worth copying, with fixtures built from work your current model gets wrong and enough repeats to report a percentage instead of an anecdote.
