Which AI model leads MathArena Apex?

Claude Opus 4.8 leads MathArena Apex with 81.3% as of 24 Oct 2026, ahead of GPT-5.5 at 80.2%.

How many models have MathArena Apex results?

46 model configurations have a MathArena Apex result on BenchLeader, all taken from MathArena.

Who runs MathArena Apex and how often is it updated?

MathArena Apex is published by MathArena. BenchLeader re-reads the published results once a day and records the date each result was published.

Does MathArena Apex count toward the BenchLeader Index?

Yes. MathArena Apex contributes to the maths category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.

MathArena Apex

The hardest problems from recent competitions. MathArena.

As of 24 Oct 2026, Claude Opus 4.8 leads MathArena Apex on BenchLeader with 81.3%, ahead of GPT-5.5 at 80.2%, across 46 model configurations with a published result.

MathArena marks 35 of these 46 results as possibly contaminated: the model was evaluated after the questions were public, so they may already be in its training data. Those scores are in the table below, and count in the index at the same weight as the rest.

Published by: MathArena
Category: Maths
Index weight: 0.5
Models: 46
Data as of: 24 Oct 2026

MIT — eth-sri/matharena.

What the test looks like

MathArena Apex collects the problems from recent competitions that no model solved at the time, forming a frontier set.

How it is scored

Percent solved, averaged over runs.

What to keep in mind

By construction scores start near zero and the set is small, so it flags breakthroughs rather than ranking ordinary ability.

Top 15

46 of 46

#
1	Claude Opus 4.8 maxAnthropic	81.3%	63.3	possible
2	GPT-5.5 xhighOpenAI	80.2%	66.1	possible
3	GPT-5.4 Pro xhighOpenAI	69.8%	61.9	possible
4	Kimi K3 thinkingMoonshot AIopen ↗	65.6%	–	possible
5	Gemini 3.1 ProGoogle	60.9%	60.6	possible
6	GPT-5.4 xhighOpenAI	54.2%	64.7	possible
7	Claude Opus 4.7 xhighAnthropic	40.6%	57.4	possible
8	Claude Opus 4.6 highAnthropic	34.5%	62.7	possible
9	Gemini 3.5 FlashGoogle	32.3%	49.1	possible
10	DeepSeek V4 Pro maxDeepSeekopen ↗	28.1%	60.8	possible
11	DeepSeek V4 Flash maxDeepSeekopen ↗	27.1%	59.4	possible
12	Gemini 3.6 FlashGoogle	26.0%	49.5	possible
13	Kimi K2.6 thinkingMoonshot AIopen ↗	24.0%	59.4	possible
14	Gemini 3 ProGoogle	23.4%	60.4	possible
15	GLM 5.2Zhipu AIopen ↗	19.8%	52.0	possible
16	Gemini 3 FlashGoogle	15.6%	57.0	possible
17	Step 3.7 FlashStepFunopen ↗	14.6%	51.5	possible
18	GPT-5.2 highOpenAI	13.5%	56.0	possible
19	Step 3.5 FlashStepFunopen ↗	13.5%	53.7	possible
20	GLM 5.1Zhipu AIopen ↗	11.5%	55.9	possible
21	GLM-5Zhipu AIopen ↗	10.9%	54.9	possible
22	DeepSeek V3.2 SpecialeDeepSeekopen ↗	9.4%	53.0	possible
23	Kimi K2.5 thinkingMoonshot AIopen ↗	8.8%	57.6	possible
24	Nemotron 3 SuperNVIDIA	7.8%	44.3	possible
25	Grok 4.1 Fast thinkingSpaceXAI	5.2%	54.7	possible
26	Grok 4 Fast RSpaceXAI	5.2%	–	possible
27	Qwen3 235B 2507 thinkingAlibaba	5.2%	–	–
28	Qwen3.5-35B-A3BAlibabaopen ↗	4.2%	50.2	possible
29	Grok 4SpaceXAI	2.1%	58.0	–
30	Qwen3.5-27BAlibabaopen ↗	2.1%	52.3	possible
31	Qwen3-4B-Instruct-2507 thinkingAlibabaopen	2.1%	40.0	–
32	GPT 5 Agent highOpenAI	2.1%	–	–
33	Claude Sonnet 4.5 thinkingAnthropic	1.6%	52.6	possible
34	Qed NanoLM-Provers	1.6%	–	possible
35	GPT-5 highOpenAI	1.0%	57.9	–
36	GPT-5.1 highOpenAI	1.0%	57.7	possible
37	GPT-5 mini highOpenAI	1.0%	54.1	–
38	GLM-4.5Zhipu AIopen ↗	1.0%	51.7	–
39	DeepSeek R1 0528DeepSeekopen ↗	1.0%	49.5	–
40	gpt-oss-120b highOpenAIopen ↗	1.0%	48.5	–
41	DeepSeek V3.2 thinkingDeepSeekopen ↗	0.5%	55.5	possible
42	Gemini 2.5 ProGoogle	0.5%	53.9	–
43	GLM-4.6Zhipu AIopen ↗	0.5%	49.9	possible
44	Qwen3 30B A3B 2507 thinkingAlibabaopen ↗	0.5%	43.2	–
45	Qwen3.5-9BAlibabaopen ↗	0.5%	40.1	possible
46	Kimi K2 Thinking thinkingMoonshot AIopen ↗	0.0%	52.4	possible

More maths benchmarks · All LLM benchmarks

Cite as: BenchLeader, “MathArena Apex leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/matharena_apex_2025.

MathArena Apex: questions

What does MathArena Apex measure?: MathArena Apex collects the problems from recent competitions that no model solved at the time, forming a frontier set. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads MathArena Apex?: Claude Opus 4.8 leads MathArena Apex with 81.3% as of 24 Oct 2026, ahead of GPT-5.5 at 80.2%.
How many models have MathArena Apex results?: 46 model configurations have a MathArena Apex result on BenchLeader, all taken from MathArena.
Who runs MathArena Apex and how often is it updated?: MathArena Apex is published by MathArena. BenchLeader re-reads the published results once a day and records the date each result was published.
Does MathArena Apex count toward the BenchLeader Index?: Yes. MathArena Apex contributes to the maths category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.