BenchLeader

Which AI benchmarks are saturated?

Data as of 15 Nov 2026 · last updated · by BenchLeader

Of the 256 benchmarks BenchLeader tracks, 32 are saturated and 39 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; 163 boards are still active and 22 are retired (no new result for a year). Maths boards saturate most: 11 of 18.

Benchmark status by categoryOf the 256 benchmarks BenchLeader tracks, 32 are saturated and 39 near-saturated: the top models score too close to the ceiling to tell apart.Benchmark status by categorySaturatedNear-saturatedActiveRetiredReasoningReasoning: 9 saturatedReasoning: 3 near-saturatedReasoning: 16 activeReasoning: 8 retired36CodingCoding: 3 saturatedCoding: 9 near-saturatedCoding: 30 activeCoding: 2 retired44Agents & toolsAgents & tools: 2 saturatedAgents & tools: 6 near-saturatedAgents & tools: 57 activeAgents & tools: 1 retired66MathsMaths: 11 saturatedMaths: 5 activeMaths: 2 retired18KnowledgeKnowledge: 5 saturatedKnowledge: 7 near-saturatedKnowledge: 17 activeKnowledge: 4 retired33Instruction followingInstruction following: 1 saturatedInstruction following: 3 near-saturatedInstruction following: 4 active8Human preferenceHuman preference: 5 activeHuman preference: 1 retired6MultimodalMultimodal: 7 near-saturatedMultimodal: 12 activeMultimodal: 4 retired23Long contextLong context: 2 near-saturatedLong context: 2 active4CompositeComposite: 2 near-saturatedComposite: 12 active14Safety & honestySafety & honesty: 1 saturatedSafety & honesty: 2 active3MetadataMetadata: 1 active1BenchLeader · data as of 15 Nov 2026 · CC BY 4.0www.benchleader.com/stats/benchmark-saturation
Of the 256 benchmarks BenchLeader tracks, 32 are saturated and 39 near-saturated: the top models score too close to the ceiling to tell apart. Hover or tap a bar for the benchmarks behind it. Chart image (SVG)

Key statistics

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

  • Of the 256 benchmarks BenchLeader tracks, 32 are saturated, 39 near-saturated, 163 active and 22 retired (BenchLeader, 15 Nov 2026). #
  • Maths has the largest share of saturated boards: 11 of 18 (BenchLeader, 15 Nov 2026). #
  • GPQA Diamond is saturated: the fifth-best model scores 95.4% (BenchLeader, 15 Nov 2026). #
  • ARC-AGI-2 is saturated: the fifth-best model scores 93.3% (BenchLeader, 15 Nov 2026). #
  • Humanity's Last Exam is still active: the leader scores 54.8% (BenchLeader, 15 Nov 2026). #
  • The hardest active boards with 10 or more models are Harvey LAB-AA v1.1 (best score 13.1%) and BTF-3 (best score 14.9%) (BenchLeader, 15 Nov 2026). #

Saturation by category

CategorySaturatedNear-saturatedActiveRetiredTotal
Reasoning9316836
Coding3930244
Agents & tools2657166
Maths11–5218
Knowledge5717433
Instruction following134–8
Human preference––516
Multimodal–712423
Long context–22–4
Composite–212–14
Safety & honesty1–2–3
Metadata––1–1
All323916322256

Saturated benchmarks (32)

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

BenchmarkCategoryFifth-bestLeaderLeader scoreModels
OTIS Mock AIMEMaths100%Gemini 4 Carbon100%289
ProofBenchMaths99%Claude Fable 5.2100%85
τ²-Bench Telecom (AA)Agents & tools98.5%GLM 5.299.1%418
ARC-AGI-1Reasoning98.5%Claude Fable 5.298.5%267
AIME 2026Maths98.3%GPT-5.5100%32
DTBenchReasoning97.6%Claude Fable 5.299.2%238
LiveBench MathematicsMaths96.7%Claude Opus 5.597.1%68
AIME (Vals)Maths96.7%Gemini 3.1 Pro98.1%94
ARC-AGI-3Reasoning96.4%GPT-6 Astra99.9%79
AIME 2025 (AA)Maths96.3%GPT-5.299%255
HMMT February 2026Maths96.2%GPT-5.598.5%32
MedQAKnowledge96.1%o196.5%93
SWE-bench (Vals)Coding95.4%Claude Opus 597%87
GPQA DiamondReasoning95.4%Gemini 4 Carbon96.5%306
MATH 500Maths95.2%Gemini 3 Pro96.4%58
FrontierMath Tier 4Maths95%Claude Fable 5.2100%79
GPQA Diamond (AA)Reasoning94.1%GPT-6 Astra96.3%567
GPQA Diamond (Vals)Reasoning93.9%Gemini 3.1 Pro95.5%134
MGSMMaths93.9%Claude Opus 4.595.2%73
ARC-AGI-2Reasoning93.3%GPT-6 Astra95%267
IFEval (HELM)Instruction following92.9%Grok 3 mini95.1%68
FACTS SearchAgents & tools92.7%Gemini 4 Carbon94.4%26
MASKSafety & honesty92.5%Claude Opus 4.696.3%65
LiveBench 2.0 ReasoningReasoning91.7%Gemini 4 Argon95.1%34
LiveBench 1.0 ReasoningReasoning91.7%GPT-6 Astra92.7%68
Global-MMLU-Lite (AA)Knowledge91.3%Gemini 3.1 Pro93.2%115
IOICoding91.2%GPT-6 Astra100%46
FrontierMath Tiers 1–3Maths90.9%Claude Fable 5.295.1%123
Vibe Code Bench v1.1Coding90.3%Claude Fable 5.293.8%126
MMLU-ProKnowledge90.2%Claude Fable 5.192.4%134
MedScribeKnowledge90.1%Claude Opus 5.591.4%108
LiveBench ResearchKnowledge90%Claude Fable 5.294.5%34

Near-saturated benchmarks (39)

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

BenchmarkCategoryFifth-bestLeaderLeader scoreModels
WeirdMLCoding89.4%GPT-6 Astra93.6%159
LiveCodeBench (AA)Coding89.4%Gemini 3 Pro91.8%307
LiveBench BrowserAgents & tools89.3%GPT-6.1 Sol93.3%34
SAGEMultimodal89.3%Claude Opus 5.591.3%29
MMMU-Pro (Vals)Multimodal89.0%Claude Fable 5.190.6%92
Fiction.LiveBench 120kLong context88.9%o3100%36
Terminal-Bench 2.1 (AA)Coding88.8%Claude Fable 5.191.4%240
LiveBench LanguageInstruction following88.7%Claude Fable 590.7%68
LiveCodeBenchCoding88.7%Claude Fable 5.190.5%141
Surface Evolver BenchCoding87.5%Claude Fable 595%29
LegalBenchKnowledge87.4%Claude Fable 588.6%148
LiveBench 2.0 CodingCoding87.1%Gemini 4 Carbon91%34
SimpleBenchReasoning86.6%Claude Fable 5.294.9%111
MMLU-Pro (HELM)Knowledge86.3%Gemini 3 Pro90.3%68
GeoBenchMultimodal86%Claude Fable 589%44
MMMU-ProMultimodal86.0%Gemini 4 Carbon89.3%296
WildBench (HELM)Instruction following85.9%Qwen3 235B A22B 250786.6%68
MCP AtlasAgents & tools85.4%Gemini 4 Argon90.7%37
AA-LCR v1.1Long context85.3%Kimi K388.7%572
Terminal-Bench 2.1 (Vals)Agents & tools84.6%Claude Opus 5.587.6%76
LiveBench 1.0 CodingCoding83.9%Claude Sonnet 5.591.4%68
Kagi LLM BenchmarkReasoning83.5%Claude Fable 591.4%137
LiveBench 2.0Composite83.4%Claude Fable 5.284.9%34
LiveBench 2.0 Data AnalysisKnowledge82.7%Gemini 4 Carbon86.5%34
AA-Omniscience v1.1: non-hallucinationKnowledge82.7%Gemini 4 Carbon95.8%142
LiveBench 1.0Composite81.6%Claude Fable 5.183.4%68
LiveBench 1.0 Data AnalysisKnowledge81.2%GPT-6 Astra83.0%68
MMMU (validation)Multimodal80.7%GPT-5.185.4%50
IFBenchInstruction following80.5%Grok 4.383.3%428
Startrise FrontendCoding80.2%Kimi K3.1 Pro82.5%37
LiveBench DesignMultimodal80.2%Claude Fable 5.285.1%34
MirrorCodeCoding78.9%Claude Sonnet 5.597.8%45
CyberGym end-to-end (AA)Agents & tools77.9%Kimi K3.1 Pro94.7%26
SimpleQA VerifiedKnowledge73.9%Claude Fable 5.294%99
Mystery Game PuzzlesReasoning73%Claude Fable 5.290%146
EBR-benchKnowledge71.4%Claude Fable 5.290.5%31
VPCTMultimodal58.7%Gemini 3 Pro91%30
CybenchAgents & tools42%Claude Opus 4.693%21
Time Horizon Index: KSPAgents & tools23.8%Claude Opus 5.591.3%13

Hardest active benchmarks

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

BenchmarkCategoryFifth-bestLeaderLeader scoreModels
Harvey LAB-AA v1.1Agents & tools8.6%Muse Wonder Full13.1%33
BTF-3Reasoning14.1%GLM 5.3 Flash14.9%13
GDPxAgents & tools9%Claude Fable 5.216.4%21
ProgramBenchCoding5.5%Claude Fable 5.219%65
CL-bench LifeReasoning16.9%GPT-5.522.2%17
SciPredictReasoning22.2%Gemini 3 Pro25.3%15
Harvey's Legal Agent Benchmark v1Agents & tools19.6%Muse Spark 1.225.4%80
Harvey's Legal Agent Benchmark v1.1Agents & tools23.3%Gemini 4 Argon27.1%37
CL-benchReasoning20.8%GPT-5.427.9%22
BrowseCompAgents & tools6.3%Gemini 3.1 Pro31.2%41

What replaced them

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

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 47 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 256 benchmarks BenchLeader tracks, 32 are saturated and 39 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 15 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 256 benchmarks BenchLeader tracks, 32 are saturated and 39 near-saturated: the top models score too close to the ceiling to tell apart. The rest: 163 active and 22 retired.
Which AI benchmarks are hardest?
Among active boards with 10 or more models, Harvey LAB-AA v1.1 (best score 13.1%) and BTF-3 (best score 14.9%).
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.

Related: all benchmarks, methodology, BenchLeader Index, AI model release statistics, all statistics.