Model benchmarks
after Anthropic's Fable 5.1 announcement table
Fable 5.1 benchmarks
7 suites · 4 models · production safeguards on
Fable 5.1Δ vs
Agentic scientific research
Terminal-Bench-Science 0.1
+23.6 pts0.0%
24.7%29.0%22.4%stderr ±3.5–4.5 pts per model
Agentic coding
Terminal-Bench 4.0
+3.5 pts0.0%
42.0%52.3%37.3%60.9% on the Mythos 5.1 harness
Knowledge work
GDPval-AA v2
+29 Elo0
172318241711
Computer use
OSWorld 2.0
+2.5 pts0.0%
72.9%75.4%—Aug 2026 task release; no comparable competitor score
Multidisciplinary reasoning
Humanity's Last Exam
+3.1 pts0.0%
57.8%56.6%—
Business workflows
AutomationBench
+4.5 pts0.0%
17.1%26.9%19.6%
Agentic coding
CursorBench 3.2.0
+2.9 pts0.0%
70.5%70.0%67.2%
Scored with production safeguards enabled: rows where a safeguard intervened score zero (OSWorld for both Fables, AutomationBench for Fable 5), which likely lowers those cells. Strips normalize per row — the printed figures are the ground truth.
after Anthropic's release tableGPT-5.6 Sol unscored on OSWorld & HLE