Benchmark Almanac

Compare models

Pick up to four models to see how they stack up on every benchmark they've all been scored on.

Scores from

Who comes out ahead?

Claude Fable 5.2 leads on 28 of the 50 benchmarks all four share. We use independent runs where they exist and fill the gaps with labs’ own numbers.

  1. Claude Fable 5.2Best on 28
  2. Gemini 4 CarbonBest on 12
  3. GPT-6 AstraBest on 7
  4. Muse WonderBest on 3

Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards

Cited by
Labs
Area
AutomationBench51.553.941.439.1 (the lab’s own number)204 of 414.8
Terminal-Bench 4.067.6 (the lab’s own number)65.2 (the lab’s own number)58.247.0174 of 420.6
LiveBench84.983.584.570.780 of 414.2
FrontierCode 2.050.242.644.025.361 of 424.9
Code ArenaElo1,8261,7681,7631,64750 of 4179
HLE-Diamond77.963.665.840.630 of 437.3
Legal Research Bench46.155.839.450.530 of 416.4
PostTrainBench 1.244.641.041.924.230 of 420.4
Blueprint-Bench 256.658.649.731.721 of 426.9
Vals Index v2.271.370.565.064.921 of 46.4
Agent Arena28.616.413.90.9520 of 427.7
Harvey LAB v1.1 (HLAB-Vals-1.1)5.0017.57.0823.820 of 418.8
Tax Agent Bench32.151.820.736.320 of 431.1
Vibe Code Bench93.889.189.683.120 of 410.7
APEX-Accounting70.470.764.767.110 of 46.0
DesignArena 3D DesignElo1,4941,4251,5361,43210 of 4111
DesignArena Game DevElo1,4021,3331,4741,41410 of 4141
DesignArena WebsiteElo1,3511,3101,3391,37710 of 467
EEBench77.383.969.362.710 of 421.2
Excel Modeling Benchmark v1.177.476.572.772.510 of 44.9
FrontierMath Tier 4 v2100.095.197.670.710 of 429.3
FrontierMath Tiers 1-3 v295.190.593.782.110 of 413.0
Humanity's Sixth Sense59.571.153.643.710 of 427.4
PRBench49.053.448.458.410 of 410.0
Terminal-Bench 4.165.864.560.347.010 of 418.8
Vals Finance Agent v2.169.567.260.465.510 of 49.0
Vals Finance Index63.667.055.661.210 of 411.4
Vals Legal Index69.275.968.475.310 of 47.5
APEX-Agents 1.187.690.475.177.000 of 415.3
AutoBaxBench88.661.459.362.900 of 429.3
BullshitBench V267.372.069.057.600 of 414.4
CAIS Risk Indexscore, lower is better8.4033.435.532.400 of 427.1
CAIS Text Capabilities Indexscore66.359.063.947.600 of 418.7
DAXBench v287.892.895.982.200 of 413.7
DesignArena Agentic OverallElo1,3691,3041,3991,17500 of 4224
DesignArena CodeElo1,3741,3411,3951,39100 of 454
DesignArena Code to VideoElo1,4401,2991,2761,21700 of 4223
DesignArena Image to HTMLElo1,2751,2171,3001,26400 of 483
DesignArena PDFElo1,3501,3241,1941,22100 of 4156
DesignArena React NativeElo1,3191,2921,2781,25700 of 462
DesignArena SpreadsheetElo1,4291,2201,3101,23200 of 4209
DesignArena UI ComponentElo1,3821,3471,3971,36400 of 450
DuelLab GameBench 2score99.294.994.377.800 of 421.4
Extended NYT Connections97.399.398.493.400 of 45.9
Furniture Assembly Benchmark90.077.580.054.200 of 435.8
GDPx16.410.99.606.8200 of 49.6
Lech Mazur Creative Writingscore3.691.753.321.7500 of 41.94
RWE-bench62.020.037.00.000 of 462.0
SWE Atlas v1.180.571.569.464.800 of 415.7
TaxCalcBench v286.075.070.084.000 of 416.0

A lab’s own number. Everything else is an independent run.Scores are percentages unless the row says otherwise. Bold is the best in each row. Click a row for its bars.

Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards

Gap is how far apart the best and worst picked models are on a test. A small gap means they're about even. A big one means the test tells them apart.

It's in points for percentages (74.3 against 68.1 is 6.2) and in the test's own unit otherwise, such as Elo. To sort mixed units together, we count those as a share of the test's full range.

Labs is how many labs report this benchmark in their launch posts and system cards, across the whole almanac. A high count means a widely used test.

Cited by is how many of the picked models' own launch posts and system cards report this benchmark. At 0, every score in the row is an independent run.