Benchmarks with data
52
Across 14 capability domains
Benchmark directory
Filter benchmarks by capability domain or evaluation independence, inspect the current leader with uncertainty bounds, or open any board to explore full results.
52
Across 14 capability domains
51
1 vendor-run
28,249
Build 17 Nov 2026
LMArena · 29 slices · 11,177 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
Gemini 4 Carbon (Google)
1,538[1,529 to 1,548]
LMArena · 29 slices · 11,177 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
Gemini 4 Carbon (Google)
1,526[1,517 to 1,536]
LMArena · 11 slices · 1,161 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
Fable 5 (Anthropic)
1,326[1,319 to 1,332]
LMArena · 11 slices · 1,161 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
Fable 5 (Anthropic)
1,309[1,303 to 1,316]
LMArena · 12 slices · 732 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
GPT Image 2.5 Sunburst (OpenAI)
1,429[1,424 to 1,434]
LMArena · 5 slices · 671 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
Fable 5.2 (Anthropic)
1,826[1,807 to 1,845]
Epoch AI Benchmarking Hub · 338 rows
GPQA Diamond, run by Epoch AI.
Gemini 4 Carbon (Google)
96.5%[94.2% to 98.9%]
Epoch AI Benchmarking Hub · 299 rows
ARC-AGI-2 results that Epoch AI republishes from the original runner.
GPT-6 Astra (OpenAI)
95%
Epoch AI Benchmarking Hub · 202 rows
Terminal-Bench results from its public leaderboard that Epoch AI republishes. Each row is a model run in an agent harness.
GPT-5.5 (OpenAI)
84.7%[80.6% to 88.8%]
Bug Hunt Bench · 145 rows
105 real bugs planted in two production repos. Coding agents find and fix them; a blind judge grades the diff against a withheld answer key.
Fable 5.2 (Anthropic)
58
LMArena · 4 slices · 119 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
GPT Image 2.5 Sunburst (OpenAI)
1,519[1,514 to 1,523]
Epoch AI Benchmarking Hub · 101 rows
FrontierMath, 2025-02-28 private set, run by Epoch AI. Superseded in benchmark_metadata by FrontierMath-Tiers-1-3-v2-Private.
GPT-5.5 Pro Pre-release (OpenAI)
52.4%[46.7% to 58.1%]
CursorBench · 91 rows
Agents are evaluated on ambiguous, multi-file tasks from real Cursor sessions. Run by Cursor (a vendor that also sells its own model, Composer).
Gemini 4 Carbon (Google)
60.4%
Epoch AI Benchmarking Hub · 72 rows
Vending-Bench 2 results that Epoch AI republishes. Scores go from about -31 to 16484, so they are not 0 to 1.
Gemini 4 Carbon (Google)
$16,484.21
LMArena · 64 rows
Arena agent leaderboard: score with confidence interval, cost and output tokens per task, observation and session counts. One row per model and category.
Fable 5.2 (Anthropic)
0.3[0.2 to 0.3]
Epoch AI Benchmarking Hub · 56 rows
Humanity's Last Exam results that Epoch AI republishes from the original runner.
GPT-6 Astra (OpenAI)
54.8%[51% to 58.6%]
Epoch AI Benchmarking Hub · 51 rows
Merge-quality coding score. Run by Cognition (a vendor); Epoch republishes the results. Aggregation per row, for example Mean@5.
Opus 5.5 (Anthropic)
54.6%
LMArena · 46 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
Fable 5.1 (Anthropic)
1,515[1,500 to 1,530]
LMArena · 46 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
Fable 5 (Anthropic)
1,500[1,493 to 1,507]
LMArena · 42 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
Qwen3.8-Max (Alibaba)
1,260[1,246 to 1,273]
LMArena · 42 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
GPT-5.6 Sol (OpenAI)
1,255[1,249 to 1,262]
Vals AI benchmarks · 38 rows
Code Migration, a component of the Vals Index.
Sonnet 5.5 (Anthropic)
87.4%[82.7% to 92.1%]
Vals AI benchmarks · 38 rows
EMB, a component of the Vals Index.
Fable 5.2 (Anthropic)
77.4%[73.1% to 81.7%]
Vals AI benchmarks · 38 rows
Finance Agent, a component of the Vals Index.
Fable 5.2 (Anthropic)
69.5%[65.6% to 73.3%]
Vals AI benchmarks · 38 rows
Legal Agent Benchmark (HLAB-Vals), a component of the Vals Index.
Gemini 4 Argon (Google)
27.1%[19.8% to 34.4%]
Vals AI benchmarks · 38 rows
Legal Research, a component of the Vals Index.
Gemini 4 Carbon (Google)
55.8%[49% to 62.5%]
Epoch AI Benchmarking Hub · 38 rows
OSWorld results from the OSWorld site and vendor posts that Epoch AI republishes. The Agent column carries the step budget, for example `(100 steps)`. Scores are already 0 to 100.
Sonnet 4.6 (Anthropic)
72.1%
Vals AI benchmarks · 38 rows
Tax Agent, a component of the Vals Index.
Gemini 4 Hydrogen (Google)
52.3%[45.2% to 59.4%]
Vals AI benchmarks · 38 rows
Terminal-Bench 4, a component of the Vals Index.
Fable 5.2 (Anthropic)
65.2%[60.6% to 69.7%]
Vals AI benchmarks · 38 rows
A single measure of AI's potential economic impact: agentic model performance across finance, coding, legal, and tax tasks, weighted by each sector's share of U.S. GDP.
Fable 5.2 (Anthropic)
71.3%[69.4% to 73.2%]
Vals AI benchmarks · 38 rows
Vibe Code Bench, a component of the Vals Index.
Fable 5.2 (Anthropic)
93.8%[91.5% to 96%]
Epoch AI Benchmarking Hub · 35 rows
SWE-bench Verified, run by Epoch AI.
Opus 4.7 (Anthropic)
83.5%[80.2% to 86.8%]
METR time horizons · 2 slices · 26 rows
Length of tasks, measured in human working time, that a model completes with 50 percent or 80 percent success. Run on METR's long-task suite.
Claude Mythos Preview (early) (Anthropic)
1,044.8 min[508.9 min to 3,304.3 min]
Epoch AI Benchmarking Hub · 11 rows
GDPval results that Epoch AI republishes. Despite the column label (%), the values are 0 to 1 fractions.
GPT-5.2 (OpenAI)
49.7%
LMArena · 2 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
Wan 2.5 Preview (Alibaba)
1,322[1,312 to 1,332]
LMArena · 2 rows
Arena human-vote leaderboard: rating with confidence interval, vote count and rank. One row per model and category.
Wan 2.6 Image (Alibaba)
1,324[1,316 to 1,331]
Artificial Analysis · 0 rows
Artificial Analysis · 0 rows
Artificial Analysis · 0 rows
Artificial Analysis · 0 rows
Artificial Analysis · 0 rows
Artificial Analysis · 0 rows
Artificial Analysis · 0 rows
Elo from human preference votes, with the 95 percent interval the source reports.
Artificial Analysis · 0 rows
Elo from human preference votes, with the 95 percent interval the source reports.
Artificial Analysis · 0 rows
Composite index of language model quality, with the cost to run it.
Artificial Analysis · 0 rows
Artificial Analysis · 0 rows
Artificial Analysis · 0 rows
Elo from human preference votes, with the 95 percent interval the source reports.
Artificial Analysis · 0 rows
Elo from human preference votes, with the 95 percent interval the source reports.
Artificial Analysis · 0 rows
Elo from human preference votes, with the 95 percent interval the source reports.
Artificial Analysis · 0 rows
Output speed and latency. Median values come from the `performance` block, mean values from `performance_v3`.
Artificial Analysis · 0 rows
List prices per million tokens.