Best AI Models for Science

Science: score by model

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

Science leaderboard
#	Model	Correct	Score	95% interval
1	GPT-5.3-CodexOpenAI	28 / 30	93.3%	79–98%
1	Claude Opus 5.5Anthropic · statistically tied with #1	28 / 30	93.3%	79–98%
3	Grok 4.7SpaceXAI · statistically tied with #1	27 / 30	90.0%	74–97%
3	GPT-6.1 SolOpenAI · statistically tied with #1	27 / 30	90.0%	74–97%
3	GPT-6 SolOpenAI · statistically tied with #1	27 / 30	90.0%	74–97%
6	GPT-6 AstraOpenAI · statistically tied with #1	26 / 30	86.7%	70–95%
6	Muse WonderMeta · statistically tied with #1	26 / 30	86.7%	70–95%
6	GPT-5.6 TerraOpenAI · statistically tied with #1	26 / 30	86.7%	70–95%
6	GPT-6.1 TerraOpenAI · statistically tied with #1	26 / 30	86.7%	70–95%
6	GPT-6.1 Terra ProOpenAI · statistically tied with #1	26 / 30	86.7%	70–95%
6	Gemini 3.1 Pro PreviewGoogle · statistically tied with #1	26 / 30	86.7%	70–95%
6	Gemini 3.8 FlashGoogle · statistically tied with #1	26 / 30	86.7%	70–95%
6	DeepSeek V4.1 Flash 1017DeepSeek · statistically tied with #1	26 / 30	86.7%	70–95%
6	Kimi K3.1 ProMoonshotAI · statistically tied with #1	26 / 30	86.7%	70–95%
15	Claude Fable 5.1Anthropic · statistically tied with #1	25 / 30	83.3%	66–93%
15	DeepSeek V4 Flash Vision ExpDeepSeek · statistically tied with #1	25 / 30	83.3%	66–93%
15	DeepSeek V4 Pro 0813DeepSeek · statistically tied with #1	25 / 30	83.3%	66–93%
15	Muse Spark OpenMeta · statistically tied with #1	25 / 30	83.3%	66–93%
15	Gemini 4 ArgonGoogle · statistically tied with #1	25 / 30	83.3%	66–93%
15	GPT-5.4OpenAI · statistically tied with #1	25 / 30	83.3%	66–93%
15	GPT-6 LunaOpenAI · statistically tied with #1	25 / 30	83.3%	66–93%
15	MiniMax M3.1 FlashMiniMax · statistically tied with #1	25 / 30	83.3%	66–93%
23	Step 5 PreviewStepFun · statistically tied with #1	24 / 30	80.0%	63–90%
23	MiniMax M3MiniMax · statistically tied with #1	24 / 30	80.0%	63–90%
23	Inkling SmallThinking Machines · statistically tied with #1	24 / 30	80.0%	63–90%
23	Qwen3.8 Max (0902)Qwen · statistically tied with #1	24 / 30	80.0%	63–90%
23	Claude Sonnet 5.5Anthropic · statistically tied with #1	24 / 30	80.0%	63–90%
28	o3OpenAI · statistically tied with #1	23 / 30	76.7%	59–88%
28	Gemini 3.5 Flash LiteGoogle · statistically tied with #1	23 / 30	76.7%	59–88%
28	Claude Haiku 5.5Anthropic · statistically tied with #1	23 / 30	76.7%	59–88%
28	gpt-oss-120bOpenAI · statistically tied with #1	23 / 30	76.7%	59–88%
32	Kimi K2.7 CodeMoonshotAI · statistically tied with #1	22 / 30	73.3%	56–86%
32	GLM 5.3 FlashZ.ai · statistically tied with #1	22 / 30	73.3%	56–86%
32	Nemotron 3 UltraNVIDIA · statistically tied with #1	22 / 30	73.3%	56–86%
32	MiMo-V2.6-ProXiaomi · statistically tied with #1	22 / 30	73.3%	56–86%
36	InklingThinking Machines · statistically tied with #1	21 / 30	70.0%	52–83%
36	Qwen3.8 2.4T A95BQwen · statistically tied with #1	21 / 30	70.0%	52–83%
36	Qwen3.8 27BQwen · statistically tied with #1	21 / 30	70.0%	52–83%
39	MiMo-V2.6-FlashXiaomi · statistically tied with #1	20 / 30	66.7%	49–81%
39	Hy3Tencent · statistically tied with #1	20 / 30	66.7%	49–81%
39	Mistral Large 4Mistral · statistically tied with #1	20 / 30	66.7%	49–81%
39	GLM 5.3Z.ai · statistically tied with #1	20 / 30	66.7%	49–81%
43	Qwen3.7 PlusQwen	19 / 30	63.3%	46–78%
43	Mercury 2.5Inception	19 / 30	63.3%	46–78%
43	Qwen3 Coder NextQwen	19 / 30	63.3%	46–78%
46	gpt-oss-20bOpenAI	18 / 30	60.0%	42–75%
46	Solar Pro 4Upstage	18 / 30	60.0%	42–75%
46	LongCat 2.0Meituan	18 / 30	60.0%	42–75%
46	Llama 4 MaverickMeta	18 / 30	60.0%	42–75%
50	Ling 3.1 FlashinclusionAI	16 / 30	53.3%	36–70%
50	Solar Mini 4Upstage	16 / 30	53.3%	36–70%
52	Nemotron 3.5 LightningNVIDIA	14 / 30	46.7%	30–64%
53	Llama 4 ScoutMeta	13 / 30	43.3%	27–61%
53	Nemotron 3.5 Super VLNVIDIA	13 / 30	43.3%	27–61%
53	Granite 4.2 8BIBM	13 / 30	43.3%	27–61%
56	Command ACohere	12 / 30	40.0%	25–58%
56	Phi 4Microsoft	12 / 30	40.0%	25–58%
58	Gemma 4 31BGoogle	11 / 30	36.7%	22–54%
58	Llama 3.3 70B InstructMeta	11 / 30	36.7%	22–54%
60	Qwen3.5 397B A17BQwen	10 / 30	33.3%	19–51%
60	Ministral 3 8B 2512Mistral	10 / 30	33.3%	19–51%
62	Qwen3.6 35B A3BQwen	9 / 30	30.0%	17–48%
63	Ministral 3 14B 2512Mistral	8 / 30	26.7%	14–44%
64	Granite 4.0 MicroIBM	7 / 30	23.3%	12–41%
64	Mistral Small 4Mistral	7 / 30	23.3%	12–41%
66	Trinity Large ThinkingArcee AI	6 / 30	20.0%	10–37%
67	Mistral Medium 3.5Mistral	5 / 30	16.7%	7–34%
67	Hermes 4 405BNous	5 / 30	16.7%	7–34%
67	Ministral 3 3B 2512Mistral	5 / 30	16.7%	7–34%
70	Gemma 4 26B A4BGoogle	4 / 30	13.3%	5–30%
71	Nova Micro 1.0Amazon	3 / 30	10.0%	3–26%
72	Reka Flash 3rekaai	0 / 30	0.0%	0–11%

How hard are these questions?

Questions by the share of 72 scored models that got them right

1No model right2 questions
2Under 25% of models2 questions
325–50% of models4 questions
450–75% of models9 questions
575% or more, not all13 questions
6Every model right0 questions

Science: frequently asked questions

Which AI model is best at science?

GPT-5.3-Codex and Claude Opus 5.5 answered the most science questions correctly: 28 of 30. With only 30 questions, 41 other models are statistically tied with the leader, so treat this ranking as a rough guide.

Where do the science questions come from?

22 from SuperGPQA (ODC-BY 1.0) and 8 from MMLU-Pro (MIT). Every question is multiple choice with ten options and one correct answer.

How hard are the science questions?

2 of the 30 questions were answered correctly by none of the 72 scored models, and 0 by all of them.

Model

Correct

Score

95% interval

GPT-5.3-CodexOpenAI

28 / 30

93.3%

79–98%

Claude Opus 5.5Anthropic · statistically tied with #1

28 / 30

93.3%

79–98%

Grok 4.7SpaceXAI · statistically tied with #1

27 / 30

90.0%

74–97%

GPT-6.1 SolOpenAI · statistically tied with #1

27 / 30

90.0%

74–97%

GPT-6 SolOpenAI · statistically tied with #1

27 / 30

90.0%

74–97%

GPT-6 AstraOpenAI · statistically tied with #1

26 / 30

86.7%

70–95%

Muse WonderMeta · statistically tied with #1

26 / 30

86.7%

70–95%

GPT-5.6 TerraOpenAI · statistically tied with #1

26 / 30

86.7%

70–95%

GPT-6.1 TerraOpenAI · statistically tied with #1

26 / 30

86.7%

70–95%

GPT-6.1 Terra ProOpenAI · statistically tied with #1

26 / 30

86.7%

70–95%

Gemini 3.1 Pro PreviewGoogle · statistically tied with #1

26 / 30

86.7%

70–95%

Gemini 3.8 FlashGoogle · statistically tied with #1

26 / 30

86.7%

70–95%

DeepSeek V4.1 Flash 1017DeepSeek · statistically tied with #1

26 / 30

86.7%

70–95%

Kimi K3.1 ProMoonshotAI · statistically tied with #1

26 / 30

86.7%

70–95%

Claude Fable 5.1Anthropic · statistically tied with #1

25 / 30

83.3%

66–93%

DeepSeek V4 Flash Vision ExpDeepSeek · statistically tied with #1

25 / 30

83.3%

66–93%

DeepSeek V4 Pro 0813DeepSeek · statistically tied with #1

25 / 30

83.3%

66–93%

Muse Spark OpenMeta · statistically tied with #1

25 / 30

83.3%

66–93%

Gemini 4 ArgonGoogle · statistically tied with #1

25 / 30

83.3%

66–93%

GPT-5.4OpenAI · statistically tied with #1

25 / 30

83.3%

66–93%

GPT-6 LunaOpenAI · statistically tied with #1

25 / 30

83.3%

66–93%

MiniMax M3.1 FlashMiniMax · statistically tied with #1

25 / 30

83.3%

66–93%

Step 5 PreviewStepFun · statistically tied with #1

24 / 30

80.0%

63–90%

MiniMax M3MiniMax · statistically tied with #1

24 / 30

80.0%

63–90%

Inkling SmallThinking Machines · statistically tied with #1

24 / 30

80.0%

63–90%

Qwen3.8 Max (0902)Qwen · statistically tied with #1

24 / 30

80.0%

63–90%

Claude Sonnet 5.5Anthropic · statistically tied with #1

24 / 30

80.0%

63–90%

o3OpenAI · statistically tied with #1

23 / 30

76.7%

59–88%

Gemini 3.5 Flash LiteGoogle · statistically tied with #1

23 / 30

76.7%

59–88%

Claude Haiku 5.5Anthropic · statistically tied with #1

23 / 30

76.7%

59–88%

gpt-oss-120bOpenAI · statistically tied with #1

23 / 30

76.7%

59–88%

Kimi K2.7 CodeMoonshotAI · statistically tied with #1

22 / 30

73.3%

56–86%

GLM 5.3 FlashZ.ai · statistically tied with #1

22 / 30

73.3%

56–86%

Nemotron 3 UltraNVIDIA · statistically tied with #1

22 / 30

73.3%

56–86%

MiMo-V2.6-ProXiaomi · statistically tied with #1

22 / 30

73.3%

56–86%

InklingThinking Machines · statistically tied with #1

21 / 30

70.0%

52–83%

Qwen3.8 2.4T A95BQwen · statistically tied with #1

21 / 30

70.0%

52–83%

Qwen3.8 27BQwen · statistically tied with #1

21 / 30

70.0%

52–83%

MiMo-V2.6-FlashXiaomi · statistically tied with #1

20 / 30

66.7%

49–81%

Hy3Tencent · statistically tied with #1

20 / 30

66.7%

49–81%

Mistral Large 4Mistral · statistically tied with #1

20 / 30

66.7%

49–81%

GLM 5.3Z.ai · statistically tied with #1

20 / 30

66.7%

49–81%

Qwen3.7 PlusQwen

19 / 30

63.3%

46–78%

Mercury 2.5Inception

19 / 30

63.3%

46–78%

Qwen3 Coder NextQwen

19 / 30

63.3%

46–78%

gpt-oss-20bOpenAI

18 / 30

60.0%

42–75%

Solar Pro 4Upstage

18 / 30

60.0%

42–75%

LongCat 2.0Meituan

18 / 30

60.0%

42–75%

Llama 4 MaverickMeta

18 / 30

60.0%

42–75%

Ling 3.1 FlashinclusionAI

16 / 30

53.3%

36–70%

Solar Mini 4Upstage

16 / 30

53.3%

36–70%

Nemotron 3.5 LightningNVIDIA

14 / 30

46.7%

30–64%

Llama 4 ScoutMeta

13 / 30

43.3%

27–61%

Nemotron 3.5 Super VLNVIDIA

13 / 30

43.3%

27–61%

Granite 4.2 8BIBM

13 / 30

43.3%

27–61%

Command ACohere

12 / 30

40.0%

25–58%

Phi 4Microsoft

12 / 30

40.0%

25–58%

Gemma 4 31BGoogle

11 / 30

36.7%

22–54%

Llama 3.3 70B InstructMeta

11 / 30

36.7%

22–54%

Qwen3.5 397B A17BQwen

10 / 30

33.3%

19–51%

Ministral 3 8B 2512Mistral

10 / 30

33.3%

19–51%

Qwen3.6 35B A3BQwen

9 / 30

30.0%

17–48%

Ministral 3 14B 2512Mistral

8 / 30

26.7%

14–44%

Granite 4.0 MicroIBM

7 / 30

23.3%

12–41%

Mistral Small 4Mistral

7 / 30

23.3%

12–41%

Trinity Large ThinkingArcee AI

6 / 30

20.0%

10–37%

Mistral Medium 3.5Mistral

5 / 30

16.7%

7–34%

Hermes 4 405BNous

5 / 30

16.7%

7–34%

Ministral 3 3B 2512Mistral

5 / 30

16.7%

7–34%

Gemma 4 26B A4BGoogle

4 / 30

13.3%

5–30%

Nova Micro 1.0Amazon

3 / 30

10.0%

3–26%

Reka Flash 3rekaai

0 / 30

0.0%

0–11%

Science: frequently asked questions

Which AI model is best at science?

Where do the science questions come from?

22 from SuperGPQA (ODC-BY 1.0) and 8 from MMLU-Pro (MIT). Every question is multiple choice with ten options and one correct answer.

How hard are the science questions?

2 of the 30 questions were answered correctly by none of the 72 scored models, and 0 by all of them.

Science: score by model

How hard are these questions?

Science: frequently asked questions

Best AI models for science

Science: score by model

How hard are these questions?

Science: frequently asked questions