AI reasoning and math benchmark leaderboard (November 2026)
As of 13 Nov 2026, BenchLeader tracks 54 reasoning and math benchmarks from 13 publishers, 24 of which feed the BenchLeader Index. Across reasoning benchmarks, GPT-6 Astra (OpenAI) leads with a reasoning score of 79.0, ahead of Claude Fable 5.2 (76.9) and Gemini 4 Carbon (74.6). Across maths benchmarks, Claude Fable 5.2 (Anthropic) leads with a maths score of 74.5, ahead of Claude Opus 5.5 (73.2) and Claude Fable 5.1 (72.0). On the most-weighted live boards, HLE-Diamond is led by Claude Fable 5.2 (max) at 80.2%; Arena Maths is led by Gemini 4 Argon (high) at 1536. Of the 54 boards, 21 are active, 4 near-saturated, 19 saturated and 10 retired (retired: no new result in a year).
Last updated · rebuilt every six hours · independent publishers only, no launch-post numbers
- benchmarks
- 54
- active
- 21
- saturated or near
- 23
- publishers
- 13
Other categories: AI agent benchmarks · AI coding benchmarks · Multimodal AI benchmarks · all LLM benchmarks
Which models lead reasoning benchmarks
Ranked on the reasoning category of the BenchLeader Index: every board below that feeds the index, standardised (50 is the average model) and averaged. One row per model, at its best reasoning effort; models need results on at least 2 of these boards.
| # | Model | Provider | Reasoning score | Index | Blended $/M |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | OpenAI | 79.0 | 66.8 | $20.00 |
| 2 | Claude Fable 5.2 (xhigh) | Anthropic | 76.9 | 69.9 | $16.00 |
| 3 | Gemini 4 Carbon (medium) | 74.6 | 65.8 | $4.00 | |
| 4 | Claude Opus 5.5 (xhigh) | Anthropic | 73.3 | 65.4 | $8.00 |
| 5 | Gemini 3.8 Flash | 73.2 | 53.9 | $1.50 | |
| 6 | Gemini 4 Hydrogen (xhigh) | 72.9 | – | $3.00 | |
| 7 | GPT-6.1 Sol (xhigh) | OpenAI | 72.1 | 63.6 | $4.00 |
| 8 | Claude Fable 5.1 (xhigh) | Anthropic | 71.9 | 64.6 | $20.00 |
| 9 | GPT-5.4 Pro | OpenAI | 70.1 | 62.8 | $67.50 |
| 10 | Claude Fable 5 (xhigh) | Anthropic | 69.6 | – | $20.00 |
Which models lead maths benchmarks
Ranked on the maths category of the BenchLeader Index: every board below that feeds the index, standardised (50 is the average model) and averaged. One row per model, at its best reasoning effort; models need results on at least 2 of these boards.
| # | Model | Provider | Maths score | Index | Blended $/M |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 (max) | Anthropic | 74.5 | 72.7 | $16.00 |
| 2 | Claude Opus 5.5 (max) | Anthropic | 73.2 | 68.4 | $8.00 |
| 3 | Claude Fable 5.1 (max) | Anthropic | 72.0 | 66.6 | $20.00 |
| 4 | Gemini 4 Carbon (high) | 71.9 | 69.5 | $4.00 | |
| 5 | GPT-6.1 Terra (max) | OpenAI | 71.7 | 64.1 | $2.00 |
| 6 | GPT-6.1 Sol (max) | OpenAI | 71.6 | 66.6 | $4.00 |
| 7 | Claude Fable 5 (max) | Anthropic | 71.4 | 65.0 | $20.00 |
| 8 | Claude Sonnet 5.5 (max) | Anthropic | 71.3 | 64.7 | $4.00 |
| 9 | GPT-6 Astra (max) | OpenAI | 70.8 | 67.1 | $20.00 |
| 10 | o3 | OpenAI | 70.6 | 60.9 | $3.50 |
Every reasoning benchmark, what it measures and who leads (36)
Reasoning benchmarks ask questions that cannot be answered by recall alone: graduate-level science (GPQA Diamond's 198 questions), expert questions written to defeat search (Humanity's Last Exam, about 2,500 of them), abstract grid puzzles (ARC-AGI) and trick questions humans find easy (SimpleBench).
Use them for: research assistants, analysis and anything where a wrong chain of thought is costly.
Caveat: MMLU is saturated and no longer separates frontier models: use MMLU-Pro, GPQA Diamond or Humanity's Last Exam instead.
| Benchmark | What it measures | Leader today | Status | Models | Newest result |
|---|---|---|---|---|---|
| HLE-Diamond (AA) | Refined 1,000-question HLE (500 reasoning, 500 knowledge) used in Intelligence Index v5. Run by Artificial Analysis. | Claude Fable 5.2 (max) 80.2% | Active | 169 | 12 Nov 2026 |
| CritPt v2 hard subset (AA) | Private 40-question hard subset of CritPt v2 used in Intelligence Index v5, with a code-execution tool… | Claude Fable 5.2 (max) 76.5% | Active | 135 | 12 Nov 2026 |
| SimpleBench | Trick questions where humans score ~84%: spatio-temporal reasoning and social intelligence. | Claude Fable 5.2 94.9% | Near-saturated | 111 | 10 Nov 2026 |
| Humanity's Last Exam | 2,500 expert-written questions across 100+ subjects (Scale AI / CAIS). | GPT-6 Astra 54.8% | Active | 48 | 7 Oct 2026 |
| Arena Hard Prompts | Text arena rating on prompts judged hard. | Gemini 4 Carbon (high) 1547 | ActiveJudge-based | 384 | 11 Nov 2026 |
| Kagi LLM Benchmark | Kagi's unpublished reasoning, coding and instruction questions. | Claude Fable 5 (thinking) 91.4% | Near-saturated | 137 | – |
| EnigmaEval | Puzzle-hunt style multi-step reasoning. Scale AI. | Claude Fable 5 (high) 39.3% | Active | 5 | 25 Mar 2026 |
| Humanity's Last Exam (AA)not in index | Artificial Analysis' own HLE run. Index v5 replaced it with HLE-Diamond; kept for history. | Claude Fable 5.2 (max) 74.3% | Active | 638 | 9 Nov 2026 |
| CritPtnot in index | Unpublished physics research problems at the level of a first PhD project. Run by Artificial Analysis… | Claude Fable 5.2 (xhigh) 35.7% | Active | 579 | 9 Nov 2026 |
| LMCAnot in index | A capability assessment with a published score ceiling. | Claude Fable 5.2 (xhigh) 70.6% | Active | 201 | 9 Nov 2026 |
| Chess Puzzlesnot in index | Solving chess puzzles from a board description. Run by Epoch AI. | Claude Fable 5.2 (max) 79.0% | Active | 157 | 10 Nov 2026 |
| Mystery Game Puzzlesnot in index | Deduction puzzles in the style of mystery games. Run by Epoch AI. | Claude Fable 5.2 (max) 90.0% | Near-saturated | 146 | 10 Nov 2026 |
| ForecastBenchnot in index | Forecasting real future events, scored as they resolve. | GPT-5.5 (high) 66.7% | Active | 89 | 3 Sep 2026 |
| GPQA Diamond (HELM)not in index | Stanford CRFM's own GPQA Diamond run. Shown for reference; Epoch's run is in the index. | Gemini 3 Pro 80.3% | ActiveContamination risk | 68 | 24 Nov 2025 |
| MysteryMechanismnot in index | Working out a hidden scientific mechanism from sparse observations. Run by Vals AI. | GPT-6 Astra (max) 53.1% | Active | 31 | 11 Nov 2026 |
| CL-benchnot in index | Continual-learning style tasks that require using information given in context. | GPT-5.4 (xhigh) 27.9% | Active | 22 | 31 Mar 2026 |
| CL-bench Lifenot in index | The everyday-life subset of CL-bench. | GPT-5.5 (high) 22.2% | Active | 17 | 24 Apr 2026 |
| SciPredictnot in index | Predicting the outcomes of scientific experiments before they are run. Scale AI. | Gemini 3 Pro 25.3% | Active | 15 | 14 Jan 2026 |
| BTF-3not in index | Forecasting tournament pooled score. | GLM 5.3 Flash 14.9% | Active | 13 | 3 Sep 2026 |
| GPQA Diamond | Graduate-level science questions written to be Google-proof. Run by Epoch AI. | Gemini 4 Carbon (high) 96.5% | SaturatedContamination risk | 306 | 10 Nov 2026 |
| ARC-AGI-2 | Semi-private ARC-AGI-2 evaluation set. | GPT-6 Astra (max) 95.0% | Saturated | 267 | 10 Nov 2026 |
| ARC-AGI-1 | Semi-private ARC-AGI-1 evaluation set. | Claude Fable 5.2 (xhigh) 98.5% | Saturated | 267 | 10 Nov 2026 |
| ARC-AGI-3 | Semi-private ARC-AGI-3 interactive evaluation set. | GPT-6 Astra (high) 99.9% | Saturated | 79 | 10 Nov 2026 |
| GPQA Diamond (AA)not in index | Artificial Analysis' own GPQA Diamond run. Shown for reference; Epoch's run is in the index. | GPT-6 Astra (xhigh) 96.3% | SaturatedContamination risk | 567 | 15 Oct 2026 |
| DTBenchnot in index | Decision-theory dilemmas scored against expert answers. | Claude Fable 5.2 99.2% | Saturated | 238 | 9 Nov 2026 |
| GPQA Diamond (Vals)not in index | Vals AI's own GPQA Diamond run. Shown for reference. | Gemini 3.1 Pro (high) 95.5% | SaturatedContamination risk | 134 | 1 Sep 2026 |
| LiveBench 1.0 Reasoningnot in index | Reasoning category of LiveBench 1.0, frozen at its final release (25 June 2026). LiveBench 2.0's Reasoning… | GPT-6 Astra (max) 92.7% | Saturated | 68 | 25 Jun 2026 |
| LiveBench 2.0 Reasoningnot in index | LiveBench 2.0 reasoning suite: short logic, spatial and theory-of-mind questions with one correct answer. | Gemini 4 Argon 95.1% | Saturated | 34 | 6 Nov 2026 |
| ARC (AI2 Challenge)not in index | Grade-school science questions, challenge set. A 2018 benchmark, saturated for current models; kept for history. | DeepSeek V3 95.3% | RetiredContamination risk | 77 | 26 Dec 2024 |
| WinoGrandenot in index | Pronoun-resolution commonsense questions. A 2019 benchmark, saturated for current models; kept for history. | Llama 3.1 405B 89.2% | RetiredContamination risk | 74 | 26 Dec 2024 |
| HellaSwagnot in index | Commonsense sentence completion. A 2019 benchmark, saturated for current models; kept for history. | GPT-4 95.3% | RetiredContamination risk | 68 | 26 Dec 2024 |
| PIQAnot in index | Physical commonsense questions. A 2019 benchmark, saturated for current models; kept for history. | GPT-4o mini 88.7% | RetiredContamination risk | 56 | 26 Dec 2024 |
| BIG-Bench Hardnot in index | Twenty-three hard tasks from BIG-Bench. A 2022 benchmark, saturated for current models; kept for history. | Gemini 1.5 Pro 001 89.2% | RetiredContamination risk | 43 | 26 Dec 2024 |
| ANLInot in index | Adversarially written natural-language inference. A 2019 benchmark, saturated for current models; kept for history. | GPT 3.5 Turbo 1106 58.1% | RetiredContamination risk | 11 | 23 Apr 2024 |
| SuperGLUEnot in index | A suite of language-understanding tasks. A 2019 benchmark, saturated for current models; kept for history. | T5-11B 88.9% | RetiredContamination risk | 8 | 27 Jan 2022 |
| CommonsenseQA 2not in index | Commonsense yes/no questions. A 2022 benchmark, saturated for current models; kept for history. | T5-11B 67.8% | RetiredContamination risk | 6 | 18 Jul 2023 |
Every maths benchmark, what it measures and who leads (18)
Maths benchmarks run from competition problems (AIME, HMMT, IMO, scored by MathArena on release day before anyone can train on them) to unpublished research-level problems (FrontierMath) and fully written proofs (ProofBench).
Use them for: quantitative analysis, scientific work and checking multi-step precision.
Caveat: GSM8K and MATH are saturated, and older AIME sets carry contamination risk because the problems are public: prefer FrontierMath and each new competition on MathArena.
| Benchmark | What it measures | Leader today | Status | Models | Newest result |
|---|---|---|---|---|---|
| Arena Maths | Text arena rating on maths prompts. | Gemini 4 Argon (high) 1536 | ActiveJudge-based | 369 | 11 Nov 2026 |
| Omni-MATH (HELM) | Olympiad-level maths problems. Run by Stanford CRFM's HELM. | GPT-5 mini 72.2% | Active | 68 | 24 Nov 2025 |
| MathArena Apex | The hardest problems from recent competitions. MathArena. | Claude Opus 4.8 (max) 81.3% | Active | 47 | – |
| IMO 2025 | International Mathematical Olympiad 2025, proof-graded. MathArena. | GPT-5 (high) 38.1% | Active | 6 | – |
| AIME 2025 (AA)not in index | All 30 problems of the 2025 American Invitational Mathematics Examination, olympiad-level with integer answers. Run by… | Nova 2.0 Lite (thinking) 94.3% | Near-saturated | 78 | 4 Jan 2026 |
| USAMO 2026not in index | USA Mathematical Olympiad 2026, proof-graded by MathArena. Shown for reference; too few models, too close to… | GPT-5.5 (xhigh) 98.2% | Active | 9 | – |
| FrontierMath Tiers 1–3 | Unpublished research-level maths problems, tiers 1–3 (v2 private set). Run by Epoch AI. | Claude Fable 5.2 (max) 95.1% | Saturated | 123 | 10 Nov 2026 |
| AIME 2026 | AIME I and II 2026, evaluated on release by MathArena. | GPT-5.5 (xhigh) 100.0% | Saturated | 32 | – |
| HMMT February 2026 | Harvard-MIT Mathematics Tournament, February 2026. MathArena. | GPT-5.5 (xhigh) 98.5% | Saturated | 32 | – |
| OTIS Mock AIME | Mock AIME competition problems from 2024–2025. Run by Epoch AI. | Gemini 4 Carbon (high) 100.0% | SaturatedContamination risk | 289 | 10 Nov 2026 |
| AIME (Vals) | Recent AIME problems, run by Vals AI at stated reasoning effort. | Gemini 3.1 Pro (high) 98.1% | SaturatedContamination risk | 94 | 16 Apr 2026 |
| ProofBench | Full written proofs graded for rigour (Vals AI). | Claude Fable 5.2 (max) 100.0% | SaturatedJudge-based | 85 | 9 Nov 2026 |
| FrontierMath Tier 4 | The hardest FrontierMath tier: problems that take expert mathematicians days. Run by Epoch AI. | Claude Fable 5.2 (max) 100.0% | Saturated | 79 | 10 Nov 2026 |
| MGSM | Grade-school maths in ten languages. Run by Vals AI. | Claude Opus 4.5 (thinking) 95.2% | SaturatedContamination risk | 73 | 9 Jan 2026 |
| LiveBench Mathematicsnot in index | Mathematics category of LiveBench 1.0, frozen at its final release (25 June 2026). LiveBench 2.0 dropped… | Claude Opus 5.5 (max) 97.1% | Saturated | 68 | 25 Jun 2026 |
| MATH 500not in index | Five hundred competition problems in algebra, probability and trigonometry. Run by Vals AI. | Gemini 3 Pro (high) 96.4% | SaturatedContamination risk | 58 | 9 Jan 2026 |
| MATH Level 5 | Hardest tier of the MATH dataset. Largely saturated, kept at half weight. Run by Epoch AI. | GPT-5 (high) 98.1% | RetiredContamination risk | 97 | 15 Oct 2025 |
| GSM8Knot in index | Grade-school maths word problems. A 2021 benchmark, saturated for current models; kept for history. | Deepseek Coder v2 94.5% | RetiredContamination risk | 73 | 29 Nov 2024 |
Reasoning and math benchmarks: questions people ask
- What are the best reasoning and math benchmarks?
- The most informative reasoning and math benchmarks today are HLE-Diamond (led by Claude Fable 5.2 (max) at 80.2%), CritPt v2 hard subset (led by Claude Fable 5.2 (max) at 76.5%) and SimpleBench (led by Claude Fable 5.2 at 94.9%): each feeds the BenchLeader Index, covers many models and has not been saturated. Use them for: research assistants, analysis and anything where a wrong chain of thought is costly. Use them for: quantitative analysis, scientific work and checking multi-step precision.
- Which AI model is best on reasoning and math benchmarks?
- As of 13 Nov 2026: Across reasoning benchmarks, GPT-6 Astra (OpenAI) leads with a reasoning score of 79.0, ahead of Claude Fable 5.2 (76.9) and Gemini 4 Carbon (74.6). Across maths benchmarks, Claude Fable 5.2 (Anthropic) leads with a maths score of 74.5, ahead of Claude Opus 5.5 (73.2) and Claude Fable 5.1 (72.0). A category score standardises every board (50 is the average model, 15 points one standard deviation) and averages them, so one board cannot decide it.
- Are reasoning and math benchmarks comparable across boards?
- Mostly, because answers are checked against a key. The caveats are contamination (public questions end up in training data, which is why fresh competitions on MathArena and private sets such as FrontierMath say more) and reasoning effort: the same model scores higher when allowed to think longer, so BenchLeader lists each effort setting separately.
- Which reasoning and math benchmarks are saturated?
- As of 13 Nov 2026, SimpleBench (near-saturated), Kagi LLM Benchmark (near-saturated), Mystery Game Puzzles (near-saturated), GPQA Diamond (saturated), ARC-AGI-2 (saturated), ARC-AGI-1 (saturated), ARC-AGI-3 (saturated) and GPQA Diamond (saturated) and 15 more. A board is saturated when its fifth-best model scores 90% or more and near-saturated when the leader reaches 90% or the fifth-best 80%; both are decided from the data and change as scores do.
- How often is this leaderboard updated?
- BenchLeader re-reads every publisher's results every six hours (runs start at 04:03, 10:03, 16:03 and 22:03 UTC), and this page is rebuilt with each update. Each board's newest result date is in the table; publishers update on their own schedules.
Related
Cite as: BenchLeader, “AI reasoning and math benchmark leaderboard (November 2026)”, data as of 13 Nov 2026, https://www.benchleader.com/benchmarks/reasoning.