Which AI model scores highest on SuperGPQA?

Claude Opus 5.5 by Anthropic, with 40 of 50 questions correct (80.0%; 95% interval 67–89%) in the leaderboard's sample of SuperGPQA.

Graduate-level multiple-choice questions across 285 disciplines, written and checked by experts. We use questions marked hard. Published by M-A-P, ByteDance Seed and collaborators under the ODC-BY 1.0 license.

How many SuperGPQA questions do you use?

50 questions (difficulty = hard, stratified by discipline), the same for every model, each with ten answer options and one correct answer.

How hard is SuperGPQA?

Of the leaderboard's 50 SuperGPQA questions, 3 were answered correctly by none of the 72 scored models and 0 by all of them.

SuperGPQA Leaderboard: AI Model Scores

SuperGPQA: score by model

50 questions · the bar shows each score's 95% interval

SuperGPQA leaderboard
#	Model	Correct	Score	95% interval
1	Claude Opus 5.5Anthropic	40 / 50	80.0%	67–89%
2	GPT-5.3-CodexOpenAI · statistically tied with #1	39 / 50	78.0%	65–87%
3	Claude Fable 5.1Anthropic · statistically tied with #1	38 / 50	76.0%	63–86%
3	Gemini 4 ArgonGoogle · statistically tied with #1	38 / 50	76.0%	63–86%
5	Grok 4.7SpaceXAI · statistically tied with #1	37 / 50	74.0%	60–84%
5	Gemini 3.8 FlashGoogle · statistically tied with #1	37 / 50	74.0%	60–84%
7	Gemini 3.1 Pro PreviewGoogle · statistically tied with #1	36 / 50	72.0%	58–83%
7	Kimi K3.1 ProMoonshotAI · statistically tied with #1	36 / 50	72.0%	58–83%
7	Claude Sonnet 5.5Anthropic · statistically tied with #1	36 / 50	72.0%	58–83%
10	GPT-6 AstraOpenAI · statistically tied with #1	35 / 50	70.0%	56–81%
10	Muse WonderMeta · statistically tied with #1	35 / 50	70.0%	56–81%
10	GPT-6.1 Terra ProOpenAI · statistically tied with #1	35 / 50	70.0%	56–81%
10	Muse Spark OpenMeta · statistically tied with #1	35 / 50	70.0%	56–81%
10	GPT-6.1 SolOpenAI · statistically tied with #1	35 / 50	70.0%	56–81%
15	GPT-5.6 TerraOpenAI · statistically tied with #1	33 / 50	66.0%	52–78%
15	GPT-6.1 TerraOpenAI · statistically tied with #1	33 / 50	66.0%	52–78%
15	Gemini 3.5 Flash LiteGoogle · statistically tied with #1	33 / 50	66.0%	52–78%
15	GPT-6 SolOpenAI · statistically tied with #1	33 / 50	66.0%	52–78%
15	GPT-6 LunaOpenAI · statistically tied with #1	33 / 50	66.0%	52–78%
20	GPT-5.4OpenAI · statistically tied with #1	32 / 50	64.0%	50–76%
20	Qwen3.8 Max (0902)Qwen · statistically tied with #1	32 / 50	64.0%	50–76%
22	DeepSeek V4 Flash Vision ExpDeepSeek · statistically tied with #1	31 / 50	62.0%	48–74%
22	DeepSeek V4.1 Flash 1017DeepSeek · statistically tied with #1	31 / 50	62.0%	48–74%
24	o3OpenAI · statistically tied with #1	30 / 50	60.0%	46–72%
24	DeepSeek V4 Pro 0813DeepSeek · statistically tied with #1	30 / 50	60.0%	46–72%
24	Claude Haiku 5.5Anthropic · statistically tied with #1	30 / 50	60.0%	46–72%
27	Inkling SmallThinking Machines · statistically tied with #1	28 / 50	56.0%	42–69%
27	Qwen3.8 27BQwen · statistically tied with #1	28 / 50	56.0%	42–69%
27	MiniMax M3.1 FlashMiniMax · statistically tied with #1	28 / 50	56.0%	42–69%
27	MiMo-V2.6-ProXiaomi · statistically tied with #1	28 / 50	56.0%	42–69%
31	InklingThinking Machines · statistically tied with #1	27 / 50	54.0%	40–67%
31	MiMo-V2.6-FlashXiaomi · statistically tied with #1	27 / 50	54.0%	40–67%
31	GLM 5.3 FlashZ.ai · statistically tied with #1	27 / 50	54.0%	40–67%
31	GLM 5.3Z.ai · statistically tied with #1	27 / 50	54.0%	40–67%
35	Step 5 PreviewStepFun	26 / 50	52.0%	39–65%
36	Kimi K2.7 CodeMoonshotAI	25 / 50	50.0%	37–63%
36	gpt-oss-120bOpenAI	25 / 50	50.0%	37–63%
38	Llama 4 MaverickMeta	23 / 49	46.9%	34–61%
39	Qwen3.8 2.4T A95BQwen	23 / 50	46.0%	33–60%
39	Qwen3 Coder NextQwen	23 / 50	46.0%	33–60%
39	MiniMax M3MiniMax	23 / 50	46.0%	33–60%
39	Nemotron 3 UltraNVIDIA	23 / 50	46.0%	33–60%
43	Hy3Tencent	22 / 50	44.0%	31–58%
44	Qwen3.7 PlusQwen	21 / 50	42.0%	29–56%
44	Mercury 2.5Inception	21 / 50	42.0%	29–56%
46	gpt-oss-20bOpenAI	18 / 50	36.0%	24–50%
47	LongCat 2.0Meituan	17 / 50	34.0%	22–48%
47	Mistral Large 4Mistral	17 / 50	34.0%	22–48%
49	Nemotron 3.5 LightningNVIDIA	16 / 50	32.0%	21–46%
50	Solar Pro 4Upstage	14 / 50	28.0%	17–42%
51	Llama 4 ScoutMeta	13 / 50	26.0%	16–40%
51	Solar Mini 4Upstage	13 / 50	26.0%	16–40%
53	Command ACohere	12 / 50	24.0%	14–37%
53	Nemotron 3.5 Super VLNVIDIA	12 / 50	24.0%	14–37%
53	Ling 3.1 FlashinclusionAI	12 / 50	24.0%	14–37%
53	Phi 4Microsoft	12 / 50	24.0%	14–37%
57	Granite 4.2 8BIBM	11 / 50	22.0%	13–35%
58	Ministral 3 8B 2512Mistral	10 / 50	20.0%	11–33%
59	Llama 3.3 70B InstructMeta	9 / 50	18.0%	10–31%
59	Qwen3.6 35B A3BQwen	9 / 50	18.0%	10–31%
61	Granite 4.0 MicroIBM	8 / 50	16.0%	8–29%
61	Ministral 3 14B 2512Mistral	8 / 50	16.0%	8–29%
63	Gemma 4 31BGoogle	7 / 50	14.0%	7–26%
63	Ministral 3 3B 2512Mistral	7 / 50	14.0%	7–26%
63	Mistral Small 4Mistral	7 / 50	14.0%	7–26%
66	Mistral Medium 3.5Mistral	5 / 50	10.0%	4–21%
66	Qwen3.5 397B A17BQwen	5 / 50	10.0%	4–21%
66	Trinity Large ThinkingArcee AI	5 / 50	10.0%	4–21%
69	Hermes 4 405BNous	3 / 50	6.0%	2–16%
69	Nova Micro 1.0Amazon	3 / 50	6.0%	2–16%
71	Reka Flash 3rekaai	1 / 50	2.0%	0–10%
71	Gemma 4 26B A4BGoogle	1 / 50	2.0%	0–10%

SuperGPQA leaderboard

SuperGPQA: score by model

How hard are these questions?

About the dataset

SuperGPQA: frequently asked questions