Logic, math & CS

Q: Which AI model is best at Logic, math & CS?

Anthropic: Claude Fable 5 leads the Logic, math & CS domain on the meo benchmark with 88% accuracy (as of 2026-10-18).

Formal logic, mathematics, and computer-science reasoning. As of 2026-10-18, the strongest model on this domain is Anthropic: Claude Fable 5 (88% — 30 of 34 questions correct).

What this domain tests

Deductive logic, mathematical problem-solving, and computer-science reasoning that require following the rules to a single correct answer rather than recalling a result.

Why it resists contamination

Problems are newly constructed on a private holdout; objective-first grading means partial-credit pattern-matching does not pass. See the full methodology.

Best AI models for Logic, math & CS

#	Model	Logic, math & CS accuracy	Questions	Overall	𝕍 index
1	Anthropic: Claude Fable 5	88%	30/34	88.7%	27.5
2	inclusionAI: Ring-2.6-1T	88%	30/34	84.7%	14.8
3	Anthropic: Claude Opus 5.5	85%	29/34	92.1%	100.0
4	Google: Gemini 3.5 Flash	85%	29/34	82.0%	11.1
5	Qwen: Qwen3 VL 32B Instruct	85%	29/34	66.1%	0.19
6	OpenAI: GPT-5.5	82%	28/34	94.1%	23.2
7	Google: Gemini 4 Argon	82%	28/34	92.5%	18.2
8	OpenAI: GPT-5.5 Pro	82%	28/34	91.6%	4.3
9	OpenAI: GPT-6.1 Sol	82%	28/34	90.4%	35.6
10	Z.ai: GLM 5.2	82%	28/34	83.2%	0.78
11	MoonshotAI: Kimi K3	82%	28/34	80.3%	3.6
12	DeepSeek: DeepSeek V4 Pro 0423	82%	28/34	80.3%	0.53
13	DeepSeek: DeepSeek V4 Flash 0423	79%	27/34	81.8%	0.43
14	Google: Gemini 3.1 Pro Preview	79%	27/34	79.5%	3.2
15	NVIDIA: Nemotron 3 Ultra	79%	27/34	75.2%	1.2
16	SpaceXAI: Grok 4.7	79%	27/34	72.4%	0.17
17	OpenAI: GPT-5.4 Nano	79%	27/34	69.5%	2.3
18	MiniMax: MiniMax M3	79%	27/34	67.4%	0.13
19	OpenAI: GPT-5 Mini	79%	27/34	66.9%	0.25
20	OpenAI: GPT-5.1-Codex-Mini	79%	27/34	66.5%	1.3
21	Xiaomi: MiMo-V2.5	79%	27/34	66.5%	0.26
22	Qwen: Qwen3 VL 235B A22B Instruct	79%	27/34	64.4%	0.060
23	Anthropic: Claude Opus 4.8	76%	26/34	84.9%	12.1
24	Qwen: Qwen3.7 Max	76%	26/34	81.0%	0.65
25	Qwen: Qwen3.6 Flash	76%	26/34	69.5%	0.17
26	MoonshotAI: Kimi K2.6	76%	26/34	66.9%	0.050
27	OpenAI: GPT-4.1 Mini	76%	26/34	63.6%	0.26
28	Qwen: Qwen3 VL 8B Thinking	76%	26/34	63.2%	0.072
29	Qwen: Qwen3.5 Plus 2026-02-15	76%	26/34	61.9%	0.012
30	ByteDance Seed: Seed 1.6 Flash	76%	26/34	57.7%	0.080
31	Qwen: Qwen3.5-Flash	74%	25/34	68.2%	0.15
32	MoonshotAI: Kimi K2.7 Code	74%	25/34	66.9%	0.073
33	SpaceXAI: Grok 4.3	74%	25/34	66.5%	1.7
34	Z.ai: GLM 4.6V	74%	25/34	65.7%	0.12
35	Owl Alpha	71%	24/34	73.7%	0.37
36	Qwen: Qwen3.7 Plus	71%	24/34	66.9%	0.051
37	Qwen: Qwen3 VL 30B A3B Instruct	71%	24/34	66.1%	0.13
38	Qwen: Qwen3 VL 30B A3B Thinking	71%	24/34	64.0%	0.008
39	Google: Gemma 4 31B	71%	24/34	63.6%	0.009
40	Mistral: Ministral 3 14B 2512	71%	24/34	56.5%	0.086
41	Xiaomi: MiMo-V2.5-Pro	68%	23/34	71.5%	0.18
42	Baidu: ERNIE 4.5 VL 424B A47B	68%	23/34	63.2%	0.015
43	Mistral: Mistral Small 4	68%	23/34	57.7%	0.065
44	Qwen: Qwen3 VL 8B Instruct	65%	22/34	56.9%	0.028
45	Mistral: Mistral Small 3.2 24B	65%	22/34	53.6%	0.017
46	Nex AGI: Nex-N2-Pro (free)	59%	20/34	65.7%	0.18
47	ByteDance Seed: Seed 1.6	59%	20/34	64.9%	0.036
48	Z.ai: GLM 5.1	59%	20/34	63.5%	0.033
49	Google: Gemini 3.1 Flash Lite Preview	59%	20/34	61.9%	0.46
50	Google: Gemini 2.5 Flash Lite	59%	20/34	57.3%	0.13
51	Mistral: Ministral 3 8B 2512	59%	20/34	54.0%	0.065
52	OpenAI: GPT-5 Nano	59%	20/34	54.0%	0.026
53	Mistral: Mistral Large 3 2512	56%	19/34	59.0%	0.17
54	StepFun: Step 3.7 Flash	56%	19/34	49.8%	0.005
55	Google: Gemini 3.1 Flash Lite	47%	16/34	56.5%	0.13
56	Tencent: Hy3 preview	47%	16/34	54.0%	0.007
57	NVIDIA: Nemotron 3 Nano Omni (free)	47%	16/34	48.5%	0.016
58	Qwen: Qwen3.6 35B A3B	47%	16/34	47.7%	0.004
59	Google: Gemma 4 26B A4B	44%	15/34	55.2%	0.003
60	Google: Gemma 3 27B	44%	15/34	50.2%	0.006
61	Perplexity: Sonar	35%	12/34	45.6%	0.004
62	Meta: Llama 4 Maverick	32%	11/34	52.7%	0.008
63	Mistral: Ministral 3 3B 2512	32%	11/34	45.6%	0.021
64	Google: Gemma 3 12B	29%	10/34	49.0%	0.009
65	Arcee AI: Trinity Large Thinking	29%	10/34	41.6%	<0.001
66	Qwen: Qwen3.5-27B	26%	9/34	49.4%	<0.001
67	OpenAI: GPT-4o-mini	26%	9/34	47.7%	0.021
68	Google: Gemma 3 4B	26%	9/34	47.3%	0.003
69	Perceptron: Perceptron Mk1	24%	8/34	57.7%	0.055
70	Qwen: Qwen2.5 VL 72B Instruct	24%	8/34	50.2%	0.005
71	NVIDIA: Nemotron Nano 12B 2 VL (free)	24%	8/34	42.3%	0.002
72	OpenAI: GPT-4.1 Nano	21%	7/34	48.1%	0.044
73	ByteDance Seed: Seed-2.0-Lite	21%	7/34	47.7%	<0.001
74	MiniMax: MiniMax-01	18%	6/34	49.4%	0.005
75	Meta: Llama 4 Scout	18%	6/34	48.5%	0.009
76	Qwen: Qwen3.6 Plus	18%	6/34	44.8%	<0.001
77	Amazon: Nova Lite 1.0	18%	6/34	39.7%	0.012
78	Qwen: Qwen3.5-35B-A3B	18%	6/34	39.3%	<0.001
79	Anthropic: Claude 3 Haiku	15%	5/34	41.8%	0.014
80	ByteDance Seed: Seed-2.0-Mini	12%	4/34	37.2%	<0.001
81	Qwen: Qwen3.5-9B	9%	3/34	41.4%	<0.001
82	Meta: Llama 3.2 11B Vision Instruct	9%	3/34	35.6%	<0.001

Primary sources & further reading

›GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in LLMs (Apple, ICLR 2025)

Frequently asked questions

What does the meo Logic, math & CS benchmark test?

Deductive logic, mathematical problem-solving, and computer-science reasoning that require following the rules to a single correct answer rather than recalling a result.

Why can't AI models game the Logic, math & CS benchmark?

Problems are newly constructed on a private holdout; objective-first grading means partial-credit pattern-matching does not pass.

Which AI model is best at Logic, math & CS?

Anthropic: Claude Fable 5 leads the Logic, math & CS domain on the meo benchmark with 88% accuracy (as of 2026-10-18).