math level 5 Benchmark: Scores and Sources

Published result; benchmark version and evaluation conditions remain in the id and result note.

pillar: math · weight 1 within pillar · unit: % (higher is better) · official board ↗

Row	Model	Result	Normalized (0–100)	Confidence
1	GPT-5 OpenAI	98.1%	98.1	100% confidence 100 percent, Full
2	GPT-5 OpenAI	97.9%	97.9	100% confidence 100 percent, Full
3	GPT-5 Mini OpenAI	97.8%	97.8	99% confidence 99 percent, High
4	o4-mini OpenAI	97.8%	97.8	87% confidence 87 percent, High
5	o3 OpenAI	97.8%	97.8	87% confidence 87 percent, High
6	Claude Sonnet 4.5 Anthropic	97.7%	97.7	80% confidence 80 percent, High
7	Qwen3 Max Alibaba / Qwen	97.1%	97.1	69% confidence 69 percent, Medium
8	GPT-5 Mini OpenAI	96.8%	96.8	99% confidence 99 percent, High
9	DeepSeek-R1 DeepSeek	96.6%	96.6	88% confidence 88 percent, High
10	Claude Haiku 4.5 Anthropic	96.4%	96.4	64% confidence 64 percent, Medium
11	GPT-5 Nano OpenAI	95.2%	95.2	85% confidence 85 percent, High
12	GPT-5 Nano OpenAI	94.9%	94.9	85% confidence 85 percent, High
13	Claude Sonnet 3.7 Anthropic	91.2%	91.2	100% confidence 100 percent, Full
14	Claude Sonnet 3.7 Anthropic	90.0%	90.0	100% confidence 100 percent, Full
15	GPT-4.1 mini OpenAI	87.3%	87.3	87% confidence 87 percent, High
16	Claude Haiku 4.5 Anthropic	86.9%	86.9	64% confidence 64 percent, Medium
17	Claude Sonnet 3.7 Anthropic	86.3%	86.3	100% confidence 100 percent, Full
18	Claude Opus 4 Anthropic	85.0%	85.0	100% confidence 100 percent, Full
19	Claude Sonnet 4 Anthropic	84.4%	84.4	100% confidence 100 percent, Full
20	GPT-4.1 OpenAI	83.0%	83.0	100% confidence 100 percent, Full
21	Mistral Medium 3 mistral	81.6%	81.6	85% confidence 85 percent, High
22	DeepSeek V3 0324 DeepSeek	75.5%	75.5	74% confidence 74 percent, Medium
23	Gemma 3 27B IT Google	74.0%	74.0	74% confidence 74 percent, Medium
24	Llama 4 Maverick 17B Instruct Meta	73.0%	73.0	78% confidence 78 percent, Medium
25	GPT-4.1 nano OpenAI	70.0%	70.0	82% confidence 82 percent, High
26	Qwen3 235B-A22B Alibaba / Qwen	68.9%	68.9	76% confidence 76 percent, Medium
27	Claude Sonnet 3.7 Anthropic	68.2%	68.2	100% confidence 100 percent, Full
28	DeepSeek-V3 DeepSeek	64.9%	64.9	93% confidence 93 percent, High
29	Llama 4 Scout 17B Instruct Meta	62.3%	62.3	64% confidence 64 percent, Medium
30	Claude Sonnet 3.5 v2 Anthropic	56.9%	56.9	100% confidence 100 percent, Full
31	Qwen Turbo Alibaba / Qwen	56.2%	56.2	47% confidence 47 percent, Low
32	GPT-4o (2024-08-06) OpenAI	53.3%	53.3	100% confidence 100 percent, Full
33	GPT-4o mini OpenAI	52.6%	52.6	100% confidence 100 percent, Full
34	GPT-4o (2024-05-13) OpenAI	51.0%	51.0	93% confidence 93 percent, High
35	Mistral Large 2.1 mistral	50.3%	50.3	100% confidence 100 percent, Full
36	GPT-4o (2024-11-20) OpenAI	49.8%	49.8	61% confidence 61 percent, Medium
37	Claude Haiku 3.5 Anthropic	46.4%	46.4	100% confidence 100 percent, Full
38	Llama-3.3-70B-Instruct Meta	41.6%	41.6	100% confidence 100 percent, Full
39	Llama-3.1-70B-Instruct Meta	36.7%	36.7	93% confidence 93 percent, High
40	Llama-3.1-8B-Instruct Meta	22.9%	22.9	93% confidence 93 percent, High
41	Claude Haiku 3 Anthropic	14.9%	14.9	93% confidence 93 percent, High

Results are as published by the source behind each value (hover or tap the number). Benchmark scores are shown individually for every source/variant (a model may have multiple rows), and vary by version, harness and date; the normalized column uses the method’s fixed 0–100 scales and feeds the math pillar of the SI Score.