基准测试类别

Reasoning 基准测试

Compare top models across the benchmark suite that best represents reasoning performance. Use this page as the fastest way to inspect the relevant tests, then jump into the full matrix when you want broader context.

类别中的基准测试

有覆盖的模型

有人类基准的基准测试

饱和的基准测试

包含内容

The current benchmark set in this category, with context on what each test captures.

所有基准测试

Graduate-Level Google-Proof Q&A (Diamond)

One of the best discriminators between models. Scores range widely (40-85%), making it highly informative for comparing reasoning ability.

accuracy %人类 65%

BIG-Bench Hard

One of the best tests of structured reasoning ability. Scores range 60-95% for frontier models, providing good differentiation.

accuracy %

AI2 Reasoning Challenge (Challenge Set)

饱和

Tests commonsense scientific reasoning. Largely saturated for frontier models but still useful for comparing mid-tier and open-source models.

accuracy %

HellaSwag Commonsense NLI

饱和

Fundamental commonsense reasoning test. Saturated for frontier models (>95%) but useful for evaluating smaller models.

accuracy %人类 95.6%

Humanity's Last Exam

The hardest academic benchmark — top models still fail 60-65% of questions. Shows how far we are from genuine expert-level reasoning.

accuracy %

Model type:

Graduate-Level Google-Proof Q&A (Diamond)

Expert-level science reasoning across biology, chemistry, and physics at PhD level. Questions are designed to be 'Google-proof' — even domain experts with web access struggle.

Metric: accuracy %Human baseline: 65%

Why it matters

One of the best discriminators between models. Scores range widely (40-85%), making it highly informative for comparing reasoning ability.

GPQA Diamond Scores (57 models)

Standard Reasoning Hybrid

LMMarketCap.com

#	Model	Provider	Type	Score	Tier
1	🥇Gemini 3.1 Pro Preview	Google	hybrid	94.3%	Elite
2	🥈Gemini 3.1 Pro	Google	hybrid	94.3%	Elite
3	🥉Claude Opus 4.7	Anthropic	hybrid	94.2%	Elite
4	Claude Fable 5	Anthropic	standard	94.1%	Elite
5	GPT-5.5	OpenAI	standard	93.6%	Elite
6	GPT-5.5 Pro	OpenAI	standard	93.6%	Elite
7	Claude Opus 4.8	Anthropic	standard	93.6%	Elite
8	Gemini 3 Pro	Google	hybrid	91.9%	Elite
9	Gemini 3 Flash	Google	hybrid	90.4%	Elite
10	DeepSeek V4 Pro	DeepSeek	standard	90.1%	Elite
11	Grok 4.3	xAI	standard	90.0%	Elite
12	GPT-5.4	OpenAI	standard	88.5%	Strong
13	Claude Opus 4.6	Anthropic	hybrid	88.4%	Strong
14	o3	OpenAI	reasoning	87.7%	Strong
15	GPT-5.2	OpenAI	standard	87.0%	Strong
16	GPT-5.1	OpenAI	standard	86.5%	Strong
17	Claude Opus 4.5	Anthropic	hybrid	86.2%	Strong
18	DeepSeek V3.2	DeepSeek	standard	85.7%	Strong
19	GPT-5	OpenAI	standard	85.0%	Strong
20	Claude 3.7 Sonnet	Anthropic	hybrid	84.8%	Strong
21	Grok 3	xAI	standard	84.6%	Strong
22	Gemma 4 31B	Google	standard	84.3%	Strong
23	Gemini 2.5 Pro	Google	hybrid	84.0%	Strong
24	Claude Sonnet 4.5	Anthropic	hybrid	83.4%	Strong
25	Gemini 2.5 Flash	Google	hybrid	82.8%	Strong
26	Claude Opus 4	Anthropic	hybrid	82.6%	Strong
27	Claude Sonnet 4.6	Anthropic	hybrid	82.0%	Strong
28	Grok 4	xAI	reasoning	82.0%	Strong
29	o4-mini	OpenAI	reasoning	81.4%	Strong
30	o3-mini	OpenAI	reasoning	79.7%	Strong
31	o1	OpenAI	reasoning	78.0%	Strong
32	o1 Preview	OpenAI	reasoning	78.0%	Strong
33	DeepSeek R1-0528	DeepSeek	reasoning	76.0%	Strong
34	Claude Sonnet 4	Anthropic	hybrid	75.4%	Strong
35	DeepSeek R1	DeepSeek	reasoning	71.5%	Good
36	Llama 4 Maverick	Meta	standard	69.8%	Good
37	Claude 3.5 Sonnet	Anthropic	standard	65.0%	Good
38	DeepSeek V3 (March 2025)	DeepSeek	standard	62.5%	Good
39	Gemini 2.0 Flash	Google	standard	62.1%	Good
40	o1-mini	OpenAI	reasoning	60.0%	Good
41	Gemini 1.5 Pro	Google	standard	59.1%	Good
42	DeepSeek V3	DeepSeek	standard	59.1%	Good
43	Llama 4 Scout	Meta	standard	57.2%	Good
44	Phi-4	Microsoft	standard	56.1%	Good
45	GPT-4.1	OpenAI	standard	56.0%	Good
46	GPT-4o	OpenAI	standard	53.6%	Good
47	Mistral Large 2	Mistral AI	standard	52.5%	Good
48	Mistral Large 2	Mistral AI	standard	52.5%	Good
49	Claude Haiku 4.5	Anthropic	standard	52.0%	Good
50	Llama 3.1 405B	Meta	standard	51.1%	Good
51	Llama 3.3 70B	Meta	standard	50.5%	Good
52	Claude 3 Opus	Anthropic	standard	50.4%	Good
53	Qwen 2.5 72B	Alibaba	standard	49.0%	Below Avg
54	GPT-4 Turbo	OpenAI	standard	48.0%	Below Avg
55	Llama 3.1 70B	Meta	standard	46.7%	Below Avg
56	Claude 3.5 Haiku	Anthropic	standard	41.6%	Below Avg
57	GPT-4o mini	OpenAI	standard	40.2%	Below Avg

How to Read This Page

Performance Tiers

Elite - Top 10% of the score range

Strong - Top 25% of the score range

Good - Above the midpoint

Below Average - Below the midpoint

Model Types

Standard - Direct inference, no chain-of-thought

Reasoning - Extended thinking (o1, R1) - slower but excels on math/reasoning

Hybrid - Optional thinking mode (Claude 3.7, Gemini 2.5) - can switch between fast and deep

Saturated benchmarks have top models clustered above 90%, making them less useful for comparison.

Scores sourced from official model cards, technical reports, and third-party evaluations (Artificial Analysis, LMSYS Arena). Last updated: 2026-08-08T12:30:08.134Z. Some scores are approximate.

Frequently Asked Questions

AI benchmarks are grouped into categories like coding, math, reasoning, knowledge, and safety. Each category contains multiple standardized tests that measure specific aspects of model performance. This page focuses on one category so you can compare models within a specific skill area.

Each benchmark has its own scoring method - accuracy percentage, pass rate, Elo rating, or normalized score. We display raw scores from official evaluations and community-run tests. Scores are updated hourly as new evaluation results become available.

A saturated benchmark is one where top models score near the maximum (typically above 95%). This means the benchmark no longer effectively differentiates between the best models, and newer, harder benchmarks are needed to measure progress.