Ranked 4 of 9 on the combined score
Claude Fable 5.1
Claude Code, max reasoning. 395 timed runs on 200 tasks from 18 benchmarks.
Duration following
Does it work for as long as it is asked? Each task was asked for a short and a long duration: the paper’s shortest and longest requests, about a quarter and four times each task’s usual time.
| Ask | Timing error | Where runs ended |
|---|---|---|
| Every timed run395 runs | 3.37× | |
| Short ask199 runs | 3.24× | |
| Long ask196 runs | 3.34× |
It ran long on short asks (median 1.77× at the short ask) and stopped early on long ones (median 0.29× at the long ask). See every run
Every Fable run
Each row is one run, with a mark on the chart. Hover a row to find its mark; click a mark to find its row.
395 runs
| Loading 395 runs. | |||
No runs match these filters.
Forecasting
Before each task, it forecast how long the task would take. Each forecast is set against its own runtime on that task: the paper’s estimate from its timed runs.
- Typical miss
- 2.47×
- Geometric mean of the factor between forecast and runtime
- Most forecasts
- 63% too high
- On 192 tasks
- Median forecast
- 15 min
- Median runtime
- 14 min
Retrospection
After its runs had finished, it was asked how long they took, with less time information at each step. Each figure is its typical miss against the real runtime, two answers per condition.
- Oracle
- 1.00×
- Native
- 1.32×
- Context-only
- 1.63×
- Replay
- 1.55×
- Scrubbed
- 1.51×
Oracle adds an elapsed-time tool. Native and Context-only continue the finished session, with tools on or off. Replay sends a rebuilt transcript to the model's API, and Scrubbed does the same without timestamps or other time cues. On 23 tasks.
Across the 18 benchmarks
Timing on its timed runs, and how good the work was. Score is the mean of its graded runs out of 100, from its middle request, about the task’s usual time; the benchmark score is the mean of these, each benchmark counting the same.
| PostTrainBench3 runs | 6.60×(few runs) | 0% | Not scored |
| METR public tasks4 runs | 5.04×(few runs) | 0% | Too few graded |
| GPQA Diamond56 runs | 4.05× | 5% | 85.728 graded |
| Agents' Last Exam24 runs | 3.71× | 0% | 57.312 graded |
| AppWorld36 runs | 3.63× | 0% | 72.218 graded |
| AssistantBench16 runs | 3.57× | 0% | 2.68 graded |
| TUA-Bench20 runs | 3.52× | 5% | 73.910 graded |
| DeepSWE24 runs | 3.49× | 0% | 83.312 graded |
| PPTArena20 runs | 3.44× | 5% | 45.010 graded |
| YC-Bench6 runs | 3.43×(few runs) | 0% | Own units |
| CORE-Bench24 runs | 3.33× | 4% | 0.012 graded |
| Terminal-Bench 4.032 runs | 3.15× | 0% | 68.816 graded |
| Humanity's Last Exam48 runs | 3.04× | 6% | 87.524 graded |
| PaperBench6 runs | 2.63×(few runs) | 0% | Not scored |
| ProgramBench24 runs | 2.62× | 0% | 67.212 graded |
| OSWorld32 runs | 2.30× | 19% | 76.216 graded |
| Sakana ALE-Bench20 runs | 1.95× | 10% | Own units |
No runs match these filters.
Highest timing error first. ALE-Bench and YC-Bench keep their own units and stay out of the score, as do PaperBench and PostTrainBench.
Read before quoting. 35 runs were stopped by AgentTime, so their time is a lower bound. Details