GPQA Diamond

Les 198 questions les plus difficiles de GPQA, où des experts titulaires d'un doctorat atteignent 65 % de précision, tandis que des non-experts qualifiés n'atteignent que 34 % malgré l'accès au web.

Modèles: 205 actuel, 617 au total
Médiane: 71.3%
Plage: 11.1%–96.3%

Distribution des scores

Le plus bas: 11.1%
Médiane: 71.3%
Le plus élevé: 96.3%

Classement

Inclure les modèles obsolètes

Rang	Modèle	Score
1	GPT-6 Astra (Xhigh) OpenAI	96.3%
2	GPT-6 Astra (Max) OpenAI	96.1%
3	Gemini 3.8 Flash (High) Google	95.3%
4	GPT-6 Astra (High) OpenAI	94.9%
5	Gemini 3.1 Pro Preview Google	94.1%
6	GPT-6 Astra (Medium) OpenAI	93.9%
7	Gemini 3.8 Flash (Medium) Google	93.5%
8	Qwen3.8 2.4T A95B Alibaba	93.5%
9	GPT-6 Astra (Low) OpenAI	93.1%
10	MiniMax-M3 MiniMax	92.9%
11	DeepSeek V4 Pro 0813 (Max) DeepSeek	92.8%
12	Qwen3.8 Max (0902) Alibaba	92.8%
13	Qwen3.8-Flash-Next Alibaba	92.3%
14	Gemini 3.8 Flash (Low) Google	92.0%
15	GLM-5.3 (Max) Z AI	91.7%
16	GPT-5.3 Codex (Xhigh) OpenAI	91.5%
17	DeepSeek V4 Flash Vision (Max) DeepSeek	91.3%
18	GLM-5.3-Flash Z AI	91.2%
19	Agnes 2.5 Pro Beta Sapiens AI	90.5%
20	Qwen3.8 27B (Xhigh) Alibaba	90.5%
21	Qwen3.7 Plus Alibaba	90.0%
22	Hy3 Tencent	89.7%
23	Kimi K2.7 Code Kimi	89.6%
24	Inkling Small Thinking Machines	89.5%
25	Qwen3.5 397B A17B (Reasoning) Alibaba	89.3%
26	Nex-N2-Pro (Based on Qwen3.5-397B-A17B) Nex AGI	89.2%
27	Solar Pro 4 Upstage	89.1%
28	Muse Spark Open (Xhigh) Meta	89.0%
29	K2 Horizon 375B A23B Institute of Foundation Models	87.3%
30	Inkling (Xhigh) Thinking Machines	87.2%
31	Nemotron 3 Ultra 550B A55B (Reasoning) NVIDIA	86.7%
32	Apodex 1.1 Apodex	86.4%
33	Ling-3.0-flash-VL InclusionAI	86.2%
34	Qwen3.5 397B A17B (Non-reasoning) Alibaba	86.1%
35	A.X-K2 SK Telecom	85.7%
36	Gemma 4 31B (Reasoning) Google	85.7%
37	Qwen3.5 122B A10B (Reasoning) Alibaba	85.7%
38	Ring-2.6-1T InclusionAI	85.7%
39	Solar Open2 250B Upstage	85.7%
40	KAT Coder Pro V2 KwaiKAT	85.5%
41	Nanbeige4.1-3B Nanbeige	84.9%
42	JT-4.1 Flash 236B A21B (Non-reasoning) China Mobile	84.5%
43	Qwen3.8 27B (Low) Alibaba	84.5%
44	Qwen3.8 27B (Medium) Alibaba	84.5%
45	Qwen3.6 35B A3B (Reasoning) Alibaba	84.1%
46	Gemini 3.5 Flash-Lite Google	83.8%
47	Muse Glimmer (High) Meta	83.5%
48	Motif 3 Motif Technologies	83.4%
49	JT-35B-Flash China Mobile	82.9%
50	K-EXAONE 2.0 LG AI Research	82.9%

1
GPT-6 Astra (Xhigh)
OpenAI
96.3%
2
GPT-6 Astra (Max)

Données de Artificial Analysis

Command Palette

GPQA Diamond

Distribution des scores

Classement