What does it mean when a benchmark is saturated?

The best models all score near the ceiling, so differences between them are noise. BenchLeader marks a percentage-scored board with 10 or more models saturated when its fifth-best model scores 90% or more, and near-saturated when the leader passes 90% or the fifth-best passes 80%.

Which AI benchmarks are hardest?

Among active boards with 10 or more models, GDPx (best score 12%) and Harvey LAB-AA v1.1 (best score 13.1%).

Which AI benchmarks are saturated?

Name: AI benchmark saturation statistics (BenchLeader)
Creator: BenchLeader
License: https://creativecommons.org/licenses/by/4.0/

Data as of 2 Nov 2026 · last updated 2 Nov 2026 · by BenchLeader

Of the 253 benchmarks BenchLeader tracks, 29 are saturated and 37 near-saturated: the top models score too close to the ceiling to tell apart. A percentage-scored board is saturated when its fifth-best model scores 90% or more; 165 boards are still active and 22 are retired (no new result for a year). Maths boards saturate most: 10 of 18.

Of the 253 benchmarks BenchLeader tracks, 29 are saturated and 37 near-saturated: the top models score too close to the ceiling to tell apart. Chart image (SVG)

Key statistics

Each statistic has its own link (#) to cite it directly.

Of the 253 benchmarks BenchLeader tracks, 29 are saturated, 37 near-saturated, 165 active and 22 retired (BenchLeader, 2 Nov 2026). #
Maths has the largest share of saturated boards: 10 of 18 (BenchLeader, 2 Nov 2026). #
GPQA Diamond is saturated: the fifth-best model scores 95.4% (BenchLeader, 2 Nov 2026). #
ARC-AGI-2 is saturated: the fifth-best model scores 91.3% (BenchLeader, 2 Nov 2026). #
Humanity's Last Exam is still active: the leader scores 54.8% (BenchLeader, 2 Nov 2026). #
The hardest active boards with 10 or more models are GDPx (best score 12%) and Harvey LAB-AA v1.1 (best score 13.1%) (BenchLeader, 2 Nov 2026). #

Saturation by category

Category	Saturated	Near-saturated	Active	Retired	Total
Reasoning	8	2	18	8	36
Coding	4	7	30	2	43
Agents & tools	1	7	55	1	64
Maths	10	1	5	2	18
Knowledge	4	6	19	4	33
Instruction following	1	3	4	–	8
Human preference	–	–	5	1	6
Multimodal	–	7	12	4	23
Long context	–	2	2	–	4
Composite	–	2	12	–	14
Safety & honesty	1	–	2	–	3
Metadata	–	–	1	–	1
All	29	37	165	22	253

Saturated benchmarks (29)

The fifth-best model scores 90% or more. Each model counts once, at its best score.

Benchmark	Category	Fifth-best	Leader	Leader score	Models
OTIS Mock AIME	Maths	100%	Claude Opus 5.5	100%	278
ProofBench	Maths	99%	Claude Opus 5.5	100%	82
τ²-Bench Telecom (AA)	Agents & tools	98.5%	GLM 5.2	99.1%	415
ARC-AGI-1	Reasoning	98.5%	GPT-6 Astra	98.5%	257
AIME 2026	Maths	98.3%	GPT-5.5	100%	31
DTBench	Reasoning	97.6%	Claude Opus 5.5	98.9%	230
LiveBench Mathematics	Maths	96.7%	Claude Opus 5.5	97.1%	68
AIME (Vals)	Maths	96.7%	Gemini 3.1 Pro	98.1%	92
HMMT February 2026	Maths	96.2%	GPT-5.5	98.5%	31
MedQA	Knowledge	96.1%	o1	96.5%	91
SWE-bench (Vals)	Coding	95.4%	Claude Opus 5	97%	86
LiveBench 2.0 Coding	Coding	95.4%	Claude Fable 5.1	97.7%	31
GPQA Diamond	Reasoning	95.4%	Gemini 4 Argon	96.4%	295
MATH 500	Maths	95.2%	Gemini 3 Pro	96.4%	57
GPQA Diamond (AA)	Reasoning	94.1%	GPT-6 Astra	96.3%	563
GPQA Diamond (Vals)	Reasoning	93.9%	Gemini 3.1 Pro	95.5%	131
MGSM	Maths	93.9%	Claude Opus 4.5	95.2%	71
IFEval (HELM)	Instruction following	92.9%	Grok 3 mini	95.1%	68
MASK	Safety & honesty	92.5%	Claude Opus 4.6	96.3%	64
LiveBench 2.0 Reasoning	Reasoning	91.7%	Gemini 4 Argon	95.1%	31
LiveBench 1.0 Reasoning	Reasoning	91.7%	GPT-6 Astra	92.7%	68
ARC-AGI-2	Reasoning	91.3%	GPT-6 Astra	95%	257
Global-MMLU-Lite (AA)	Knowledge	91.3%	Gemini 3.1 Pro	93.2%	32
IOI	Coding	91.2%	GPT-6 Astra	100%	46
Vibe Code Bench v1.1	Coding	90.3%	Claude Sonnet 5.5	92.4%	121
MMLU-Pro	Knowledge	90.2%	Claude Fable 5.1	92.4%	131
FrontierMath Tier 4	Maths	90.2%	GPT-6.1 Sol	100%	76
FrontierMath Tiers 1–3	Maths	90.2%	GPT-6 Astra	93.7%	120
MedScribe	Knowledge	90.1%	Claude Opus 5.5	91.4%	107

Near-saturated benchmarks (37)

The leader is at 90% or more, or the fifth-best model at 80% or more.

Benchmark	Category	Fifth-best	Leader	Leader score	Models
WeirdML	Coding	89.4%	GPT-6 Astra	93.6%	157
LiveCodeBench (AA)	Coding	89.4%	Gemini 3 Pro	91.8%	304
AIME 2025 (AA)	Maths	89.3%	Nova 2.0 Lite	94.3%	78
LiveBench Browser	Agents & tools	89.3%	GPT-6.1 Sol	93.3%	31
SAGE	Multimodal	89.3%	Claude Opus 5.5	91.3%	28
MMMU-Pro (Vals)	Multimodal	89.0%	Claude Fable 5.1	90.6%	91
Fiction.LiveBench 120k	Long context	88.9%	o3	100%	35
Terminal-Bench 2.1 (AA)	Coding	88.8%	Claude Fable 5.1	91.4%	238
LiveBench Language	Instruction following	88.7%	Claude Fable 5	90.7%	68
LiveCodeBench	Coding	88.7%	Claude Fable 5.1	90.5%	138
Surface Evolver Bench	Coding	87.5%	Claude Fable 5	95%	29
LegalBench	Knowledge	87.4%	Claude Fable 5	88.6%	145
MMLU-Pro (HELM)	Knowledge	86.3%	Gemini 3 Pro	90.3%	68
FACTS Search	Agents & tools	86.2%	GPT-6 Astra	89.9%	22
WildBench (HELM)	Instruction following	85.9%	Qwen3 235B A22B 2507	86.6%	68
MMMU-Pro	Multimodal	85.6%	Gemini 4 Carbon	89.3%	292
AA-LCR v1.1	Long context	85.3%	Kimi K3	88.7%	560
Terminal-Bench 2.1 (Vals)	Agents & tools	84.6%	Claude Opus 5.5	87.6%	75
LiveBench Research	Knowledge	84.5%	Muse Wonder	88%	31
MCP Atlas	Agents & tools	84.2%	Muse Spark 1.1	88.1%	34
LiveBench 1.0 Coding	Coding	83.9%	Claude Sonnet 5.5	91.4%	68
Kagi LLM Benchmark	Reasoning	83.5%	Claude Fable 5	91.4%	136
LiveBench Design	Multimodal	83.1%	Claude Opus 5.5	86.9%	31
AA-Omniscience v1.1: non-hallucination	Knowledge	82.0%	Gemini 4 Carbon	95.8%	110
LiveBench 1.0	Composite	81.6%	Claude Fable 5.1	83.4%	68
LiveBench 2.0 Data Analysis	Knowledge	81.6%	Gemini 4 Argon	85.2%	31
LiveBench 1.0 Data Analysis	Knowledge	81.2%	GPT-6 Astra	83.0%	68
MMMU (validation)	Multimodal	80.7%	GPT-5.1	85.4%	50
IFBench	Instruction following	80.5%	Grok 4.3	83.3%	425
Startrise Frontend	Coding	80.2%	Kimi K3.1 Pro	82.5%	37
LiveBench 2.0	Composite	80.0%	Claude Fable 5.1	85.7%	31
GeoBench	Multimodal	80%	Gemini 3 Flash	88%	27
CyberGym end-to-end (AA)	Agents & tools	76.3%	Kimi K3.1 Pro	94.7%	22
VPCT	Multimodal	58.7%	Gemini 3 Pro	91%	30
Cybench	Agents & tools	42%	Claude Opus 4.6	93%	21
ARC-AGI-3	Reasoning	35%	GPT-6 Astra	99.9%	70
Time Horizon Index: KSP	Agents & tools	23.8%	Claude Opus 5.5	91.3%	13

Hardest active benchmarks

Active percentage-scored boards with results for 10 or more models, lowest leading score first.

Benchmark	Category	Fifth-best	Leader	Leader score	Models
GDPx	Agents & tools	8.5%	Claude Opus 5.5	12%	20
Harvey LAB-AA v1.1	Agents & tools	8.6%	Muse Wonder Full	13.1%	28
BTF-3	Reasoning	14.1%	GLM 5.3 Flash	14.9%	13
ProgramBench	Coding	3%	Claude Opus 5.5	18.5%	63
Remote Labor Index	Agents & tools	16.1%	Claude Opus 5.5	21.3%	17
CL-bench Life	Reasoning	16.9%	GPT-5.5	22.2%	16
SciPredict	Reasoning	22.2%	Gemini 3 Pro	25.3%	15
Harvey's Legal Agent Benchmark v1	Agents & tools	19.6%	Muse Spark 1.2	25.4%	79
Harvey's Legal Agent Benchmark v1.1	Agents & tools	18.3%	Gemini 4 Argon	27.1%	33
CL-bench	Reasoning	20.8%	GPT-5.4	27.9%	21

What replaced them

For saturated and retired boards, the live benchmarks that test the same skill and still separate the top models.

GPQA Diamond (AA) (saturated) → Humanity's Last Exam, CritPt
τ²-Bench Telecom (AA) (saturated) → τ³-Banking (AA), Terminal-Bench
GPQA Diamond (saturated) → Humanity's Last Exam, CritPt
OTIS Mock AIME (saturated) → AIME 2026, FrontierMath Tiers 1–3
ARC-AGI-1 (saturated) → ARC-AGI-2, ARC-AGI-3
GPQA Diamond (Vals) (saturated) → Humanity's Last Exam, CritPt
AIME (Vals) (saturated) → AIME 2026, FrontierMath Tiers 1–3
MedQA (saturated) → MedScribe, AA-Omniscience v1.1
MGSM (saturated) → FrontierMath Tiers 1–3, AIME 2026
IFEval (HELM) (saturated) → IFBench, MultiChallenge
MATH 500 (saturated) → FrontierMath Tiers 1–3, AIME 2026
MMLU (retired) → MMLU-Pro, GPQA Diamond, Humanity's Last Exam
MATH Level 5 (retired) → FrontierMath Tiers 1–3, HMMT February 2026
ARC (AI2 Challenge) (retired) → ARC-AGI-2, GPQA Diamond
WinoGrande (retired) → SimpleBench, Humanity's Last Exam

How these numbers are made

BenchLeader judges saturation only on percent-scored benchmarks with results for at least 10 models. Saturated: the fifth-best model (one score per model, its best) is at 90% or more, so the leaders are separated by noise. Near-saturated: the leader is at 90% or more, or the fifth-best at 80% or more. Retired: no new result for more than 365 days. Other benchmarks (Elo ratings, dollar amounts, small boards) are ‘active’ unless retired. Counts are benchmarks BenchLeader tracks from 45 publishers, not every benchmark published.

Some trackers call a benchmark saturated when the top model passes 90%. BenchLeader uses the fifth-best model, so one outlier result cannot saturate a board. The statuses are the same ones the benchmark catalogue shows.

Use this chart

These counts are BenchLeader's own and free to reuse under CC BY 4.0: credit BenchLeader with a link to this page. The benchmark scores in the tables and CSV remain under each benchmark publisher's own terms.

Embed code

<figure><a href="https://www.benchleader.com/stats/benchmark-saturation"><img src="https://www.benchleader.com/stats/benchmark-saturation.svg" alt="Of the 253 benchmarks BenchLeader tracks, 29 are saturated and 37 near-saturated: the top models score too close to the ceiling to tell apart." width="760"></a><figcaption>Source: <a href="https://www.benchleader.com/stats/benchmark-saturation">BenchLeader</a>, CC BY 4.0</figcaption></figure>

Download the table (CSV)

Citation

BenchLeader, “AI benchmark saturation statistics”, data as of 2 Nov 2026, https://www.benchleader.com/stats/benchmark-saturation.

Questions people ask

What does it mean when a benchmark is saturated?: The best models all score near the ceiling, so differences between them are noise. BenchLeader marks a percentage-scored board with 10 or more models saturated when its fifth-best model scores 90% or more, and near-saturated when the leader passes 90% or the fifth-best passes 80%.
How many AI benchmarks are saturated?: Of the 253 benchmarks BenchLeader tracks, 29 are saturated and 37 near-saturated: the top models score too close to the ceiling to tell apart. The rest: 165 active and 22 retired.
Which AI benchmarks are hardest?: Among active boards with 10 or more models, GDPx (best score 12%) and Harvey LAB-AA v1.1 (best score 13.1%).
Why use the fifth-best model rather than the top score?: Some trackers call a benchmark saturated when the top model passes 90%. BenchLeader uses the fifth-best model, so one outlier result cannot saturate a board.