Acropolis

Fable 5.2

AnthropicGrade A. Evidence grade A of A to D: strongest.ProprietaryReleased 9 Nov 2026Consistent across benchmarks
Compare with peers

Domain profile

Median within-board percentile by capability domain across boards with sufficient peer overlap.

  • Agentic (12 boards): 99 (scale 0 to 100)Agentic (12 boards)99
  • Coding (8 boards): 100 (scale 0 to 100)Coding (8 boards)100
  • Finance (4 boards): 100 (scale 0 to 100)Finance (4 boards)100
  • General text (1 boards): 97 (scale 0 to 100)General text (1 boards)97
  • Legal (3 boards): 65 (scale 0 to 100)Legal (3 boards)65
  • Science (2 boards): 98 (scale 0 to 100)Science (2 boards)98

Benchmark appearances

Best verified run per board, with confidence intervals and Pareto frontier status.

BenchmarkRankScoreCostStatusObserved
Bug Hunt Bench

Bug Hunt Bench · max · Claude Code

#1 / 145Tied #1
58$139.65On frontier. Evidence grade A of A to D: strongest.9 Nov 2026
EMB

Vals AI benchmarks · high

#1 / 38Tied #1
77.4%[73.1% to 81.7%]$6.67On frontier. Evidence grade A of A to D: strongest.15 Nov 2026
Finance Agent

Vals AI benchmarks · high

#1 / 38Tied #1
69.5%[65.6% to 73.3%]$4.83On frontier. Evidence grade A of A to D: strongest.15 Nov 2026
LMArena Agent

LMArena · Overall · max

#1 / 64Tied #1
0.3[0.2 to 0.3]$9.02On frontier. Evidence grade A of A to D: strongest.15 Nov 2026
LMArena WebDev

LMArena · Overall · xhigh

#1 / 156Tied #1
1,826[1,807 to 1,845]UnknownStandard15 Nov 2026
LMArena WebDev

LMArena · Webdev · xhigh

#1 / 156Tied #1
1,826[1,807 to 1,845]UnknownStandard15 Nov 2026
LMArena WebDev

LMArena · Webdev react · xhigh

#1 / 140Tied #1
1,895[1,869 to 1,921]UnknownStandard15 Nov 2026
Terminal-Bench 4

Vals AI benchmarks · high

#1 / 38Tied #1
65.2%[60.6% to 69.7%]$14.56Standard15 Nov 2026
Vals Index

Vals AI benchmarks · high

#1 / 38Tied #1
71.3%[69.4% to 73.2%]$17.94On frontier. Evidence grade A of A to D: strongest.15 Nov 2026
Vibe Code Bench

Vals AI benchmarks · high · OpenHands

#1 / 38Tied #1
93.8%[91.5% to 96%]$22.13On frontier. Evidence grade A of A to D: strongest.15 Nov 2026
Code Migration

Vals AI benchmarks · high

#2 / 38Tied #1
85.8%[78.9% to 92.8%]$68.68Standard15 Nov 2026
CursorBench

CursorBench · max · Cursor

#3 / 91
60%$32.79StandardNot stated
GPQA diamond

Epoch AI Benchmarking Hub · max · fallback

#3 / 338Tied #1
96%[93.4% to 98.6%]UnknownStandard6 Nov 2026
ARC-AGI-2

Epoch AI Benchmarking Hub · xhigh

#4 / 299
93.8%UnknownStandardNot stated
LMArena WebDev

LMArena · Webdev HTML · xhigh

#4 / 155Tied #1
1,760[1,717 to 1,804]UnknownStandard15 Nov 2026
Bug Hunt Bench

Bug Hunt Bench · xhigh · Claude Code

#5 / 145
49$61.01Standard10 Nov 2026
ARC-AGI-2

Epoch AI Benchmarking Hub · max

#7 / 299
92.5%UnknownStandardNot stated
CursorBench

CursorBench · high · Cursor

#7 / 91
56.9%$9.70StandardNot stated
CursorBench

CursorBench · xhigh · Cursor

#8 / 91
56.3%$15.48StandardNot stated
Bug Hunt Bench

Bug Hunt Bench · high · Claude Code

#10 / 145
43$46.67Standard10 Nov 2026
CursorBench

CursorBench · medium · Cursor

#12 / 91
54.8%$4.66StandardNot stated
Legal Research

Vals AI benchmarks · high

#13 / 38
46.2%[39.4% to 52.9%]$3.96Standard15 Nov 2026
CursorBench

CursorBench · low · Cursor

#14 / 91
53.2%$3.94On frontier. Evidence grade A of A to D: strongest.Not stated
ARC-AGI-2

Epoch AI Benchmarking Hub · high

#18 / 299
90.4%UnknownStandardNot stated
Bug Hunt Bench

Bug Hunt Bench · medium · Claude Code

#18 / 145
41.5$26.55Standard10 Nov 2026
Bug Hunt Bench

Bug Hunt Bench · low · Claude Code

#20 / 145
41$24.32Standard10 Nov 2026
GPQA diamond

Epoch AI Benchmarking Hub · max

#20 / 338Tied #1
93.9%[90.7% to 97.1%]UnknownStandard6 Nov 2026
Tax Agent

Vals AI benchmarks · high

#21 / 38
32.1%[25.5% to 38.7%]$3.31Standard15 Nov 2026
ARC-AGI-2

Epoch AI Benchmarking Hub · medium

#23 / 299
89.6%UnknownStandardNot stated
Legal Agent Benchmark (HLAB-Vals)

Vals AI benchmarks · high

#30 / 38
5%[1.6% to 8.4%]$19.37Standard15 Nov 2026
ARC-AGI-2

Epoch AI Benchmarking Hub · low

#34 / 299
86.7%UnknownStandardNot stated

Reasoning effort scaling

Same benchmark and harness measured across multiple reasoning effort settings.

  • ARC-AGI-2

    Epoch AI Benchmarking Hub

    1 regression. Evidence grade D of A to D: weakest.
    low

    86.7

    medium

    89.6

    high

    90.4

    xhigh

    93.8

    max

    92.5

  • Bug Hunt Bench

    Bug Hunt Bench · Claude Code

    Monotone. Evidence grade A of A to D: strongest.
    low

    41.0

    $24.32

    medium

    41.5

    $26.55

    high

    43.0

    $46.67

    xhigh

    49.0

    $61.01

    max

    58.0

    $139.65

  • CursorBench

    CursorBench · Cursor

    1 regression. Evidence grade D of A to D: weakest.
    low

    53.2

    $3.94

    medium

    54.8

    $4.66

    high

    56.9

    $9.70

    xhigh

    56.3

    $15.48

    max

    60.0

    $32.79

Closest head-to-head rivals

Direct pairwise win rates for Fable 5.2 across shared benchmark boards.