The evidence layer for AI model claims

Benchmark Integrity Index

A score is f(model, harness, protocol), not f(model). We evaluate the evaluations — grading each benchmark on discrimination, saturation, contamination, and harness comparability — so you know which numbers still mean something.

61 benchmarks graded, most trustworthy first. Every input is source-backed.

ABCDdot size = models scored · 61 benchmarks

Grade

Harness

Sort

#	Grade	Benchmark	Signals	Top score
1	A100	ARC-AGI-2	consistent harnessheld-out test set	GPT-5.6 Sol 92.5 · 71 models
2	A100	Chess Puzzles	consistent harnessheld-out test set	GPT-5.5 Pro 62.12 · 81 models
3	A100	HLE	consistent harnesspublic test set	Gemini 3.1 Pro 43.74 · 37 models
4	A100	SimpleQA Verified	consistent harnesspublic test set	Gemini 3.1 Pro 77.3 · 68 models
5	A99	APEX-Agents	consistent harnesspublic test set	Claude Fable 5 45 · 47 models
6	A99	EBR-bench	consistent harnesspublic test set	Claude Opus 5 50 · 18 models
7	A99	FrontierCode	consistent harnesspublic test set	Claude Fable 5 53.5 · 25 models
8	A99	FrontierMath-2025-02-28-Private	consistent harnessheld-out test set	Muse Spark 68.42 · 45 models
9	A99	FrontierMath-Tier-4-v2-Private	consistent harnessheld-out test set	Claude Fable 5 87.8 · 48 models
10	A99	FrontierMath-Tiers-1-3-v2-Private	consistent harnessheld-out test set	GPT-5.6 Sol 89.12 · 48 models
11	A99	Mystery Game Puzzles	consistent harnesspublic test set	Claude Opus 5 54.84 · 35 models
12	A99	ProofBench	consistent harnesspublic test set	Claude Opus 5 99 · 51 models
13	A98	CadEval	consistent harnesspublic test set	o3 74 · 14 models
14	A98	DeepSWE	consistent harnesspublic test set	Claude Opus 5 73.65 · 21 models
15	A98	GDPval	consistent harnessheld-out test set	GPT-5.2 49.7 · 11 models
16	A98	LiveCodeBench	consistent harnesspublic test set	DeepSeek-R1 65.9 · 8 models
17	A98	SimpleBench	consistent harnessheld-out test set	Claude Fable 5 78.28 · 71 models
18	A97	AIME	consistent harnesspublic test set	DeepSeek-R1 79.8 · 8 models
19	A97	MirrorCode	consistent harnesspublic test set	Claude Fable 5 63.89 · 6 models
20	A96	Codeforces rating	consistent harnesspublic test set	DeepSeek-R1 2029 · 8 models
21	A96	Surface Evolver Bench	consistent harnesspublic test setsaturated	Kimi K3 95 · 19 models
22	A95	PostTrainBench	consistent harnesspublic test set	Claude Fable 5 41.79 · 25 models
23	A94	CritPt	consistent harnessheld-out test set	GPT-5.6 Sol 32.3 · 76 models
24	A94	ExploitBench	consistent harnesspublic test set	GPT-5.5 47.4 · 8 models
25	A93	FrontierMath-Tier-4-2025-07-01-Private	consistent harnessheld-out test set	Gemini 3 Pro 31.25 · 27 models
26	A91	METR Time Horizons	mixed harness — not comparablepublic test set	Claude Opus 4.6 78.86 · 37 models
27	A91	The Agent Company	consistent harnesspublic test set	DeepSeek-V3.2-Exp 42.9 · 14 models
28	A89	Balrog	consistent harnesspublic test set	Gemini 3 Pro 58.1 · 24 models
29	A89	GSO-Bench	consistent harnesspublic test set	Claude Opus 4.8 47.06 · 24 models
30	A88	GeoBench	consistent harnesspublic test set	Gemini 3 Flash 88 · 26 models
31	A87	Aider polyglot	consistent harnesspublic test set	GPT-5 88 · 43 models
32	A87	WeirdML	consistent harnessheld-out test set	Claude Fable 5 91.94 · 108 models
33	A86	Fiction.LiveBench	consistent harnesspublic test set	o3-pro 97.2 · 43 models
34	A86	OTIS Mock AIME 2024-2025	consistent harnesspublic test setsaturated	Claude Fable 5 100 · 136 models
35	A86	VPCT	consistent harnesspublic test set	Gemini 3 Pro 86.5 · 26 models
36	B84	Terminal Bench	consistent harnesspublic test set	GPT-5.5 84.7 · 36 models
37	B83	SWE-Bench verified	consistent harnesspublic test set	Claude Opus 4.7 83.47 · 32 models
38	B82	CL-bench	consistent harnesspublic test set	GPT-5.4 27.9 · 19 models
39	B82	OSWorld 2.0	consistent harnesspublic test set	Claude Opus 4.8 20.6 · 7 models
40	B81	CL-bench Life	consistent harnesspublic test set	GPT-5.5 22.2 · 13 models
41	B81	Remote Labor Index	consistent harnesspublic test set	Claude Fable 5 16.1 · 10 models
42	B80	OSWorld	consistent harnesspublic test set	Claude Sonnet 4.6 72.1 · 8 models
43	B79	Cybench	mixed harness — not comparablepublic test set	Claude Opus 4.6 93 · 19 models
44	B77	DeepResearch Bench	consistent harnesspublic test set	Claude Opus 4.6 55.31 · 22 models
45	B76	Lech Mazur Writing	consistent harnesspublic test set	GPT-5 86 · 41 models
46	B76	MATH-500	consistent harnesspublic test set	DeepSeek-R1 97.3 · 8 models
47	B73	ARC-AGI	consistent harnesspublic test setsaturated	Claude Fable 5 98.5 · 72 models
48	B73	GPQA diamond	mixed harness — not comparablepublic test set	Gemini 3.7 Flash 93.1 · 153 models
49	B73	MATH level 5	consistent harnesspublic test setsaturated	GPT-5 98.13 · 79 models
50	B71	BBH	mixed harness — not comparablepublic test set	Gemini 1.5 Pro (May 2024) 85.6 · 41 models
51	C68	GSM8K	mixed harness — not comparablepublic test set	GPT-4 (Mar 2023) 92 · 56 models
52	C67	HellaSwag	mixed harness — not comparablepublic test set	GPT-4 (Mar 2023) 93.73 · 52 models
53	C67	MMLU	mixed harness — not comparablepublic test set	GPT-4o (Nov 2024) 84.13 · 99 models
54	C67	ScienceQA	mixed harness — not comparablepublic test set	GPT-4o (May 2024) 84.67 · 6 models
55	C66	ARC AI2	mixed harness — not comparablepublic test set	DeepSeek-V3 93.73 · 66 models
56	C63	TriviaQA	mixed harness — not comparablepublic test set	Llama 2-70B 87.6 · 32 models
57	C62	PIQA	mixed harness — not comparablepublic test set	PowerMoE-3b 79.1 · 44 models
58	C62	Winogrande	mixed harness — not comparablepublic test set	Llama 3.1-405B 78.4 · 64 models
59	C61	OpenBookQA	mixed harness — not comparablepublic test set	phi-3-mini 3.8B 84 · 33 models
60	C59	ANLI	consistent harnesspublic test set	phi-3-small 7.4B 37.15 · 9 models
61	D54	LAMBADA	mixed harness — not comparablepublic test set	Falcon-180B 79.8 · 20 models

How the score works

Integrity = 0.30·discrimination + 0.30·saturation + 0.15·contamination + 0.15·harness + 0.10·age. A high grade means the benchmark still separates models, is not near ceiling, resists training-set contamination, and is scored under a consistent harness. It does not mean the number reflects real-world or agentic performance. That is a separate question we flag but never conflate. Full methodology →