Which AI model leads AIME 2026?

GPT-5.5 leads AIME 2026 with 100.0% as of 25 Oct 2026, ahead of Claude Opus 4.8 at 100.0%.

How many models have AIME 2026 results?

31 model configurations have a AIME 2026 result on BenchLeader, all taken from MathArena.

Who runs AIME 2026 and how often is it updated?

AIME 2026 is published by MathArena. BenchLeader re-reads the published results once a day and records the date each result was published.

Does AIME 2026 count toward the BenchLeader Index?

Yes. AIME 2026 contributes to the maths category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.

AIME 2026

AIME I and II 2026, evaluated on release by MathArena.

As of 25 Oct 2026, GPT-5.5 leads AIME 2026 on BenchLeader with 100.0%, ahead of Claude Opus 4.8 at 100.0%, across 31 model configurations with a published result.

MathArena marks 22 of these 31 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: 1.0
Models: 31
Data as of: 25 Oct 2026

MIT — eth-sri/matharena.

What the test looks like

The 2026 AIME I and II papers, evaluated by MathArena within days of the exam so that no model could have trained on them.

How it is scored

Percent of the 30 problems answered correctly, averaged over several runs.

What to keep in mind

Thirty problems, so each is worth over three points; run-to-run variance is published and can be several points.

Top 15

31 of 31

#
1	GPT-5.5 xhighOpenAI	100.0%	66.1	possible
2	Claude Opus 4.8 maxAnthropic	100.0%	63.3	possible
3	GPT-5.4 xhighOpenAI	99.2%	64.7	possible
4	Gemini 3.1 ProGoogle	98.3%	60.5	possible
5	GPT-5.2 highOpenAI	98.3%	56.0	–
6	Claude Opus 4.6 highAnthropic	96.7%	62.7	–
7	DeepSeek V4 Pro maxDeepSeekopen ↗	96.7%	60.8	possible
8	Gemini 3 FlashGoogle	96.7%	57.0	–
9	GLM-5Zhipu AIopen ↗	96.7%	54.9	possible
10	Step 3.5 FlashStepFunopen ↗	96.7%	53.7	–
11	Gemini 3.6 FlashGoogle	96.7%	49.5	possible
12	Kimi K3 thinkingMoonshot AIopen ↗	96.7%	–	possible
13	DeepSeek V4 Flash maxDeepSeekopen ↗	95.8%	59.4	possible
14	Kimi K2.6 thinkingMoonshot AIopen ↗	95.8%	59.4	possible
15	Kimi K2.5 thinkingMoonshot AIopen ↗	95.8%	57.6	–
16	Claude Opus 4.7 xhighAnthropic	95.8%	57.4	possible
17	GLM 5.1Zhipu AIopen ↗	95.8%	55.9	possible
18	Step 3.7 FlashStepFunopen ↗	95.0%	51.5	possible
19	Gemini 3.5 FlashGoogle	95.0%	48.8	possible
20	Grok 4.1 Fast thinkingSpaceXAI	94.2%	54.7	–
21	Qwen3.5 397B A17BAlibabaopen ↗	94.2%	51.5	possible
22	Qwen3.5-35B-A3BAlibabaopen ↗	93.3%	50.2	possible
23	Qwen3.5-9BAlibabaopen ↗	92.5%	40.1	possible
24	Gemini 3 ProGoogle	91.7%	60.4	–
25	Qwen3.5-27BAlibabaopen ↗	91.7%	52.3	possible
26	Nemotron 3 SuperNVIDIA	91.7%	44.3	possible
27	GLM 5.2Zhipu AIopen ↗	90.0%	51.8	possible
28	Qwen3.5-4BAlibabaopen ↗	89.7%	–	possible
29	Qwen3 30B A3B 2507 thinkingAlibabaopen ↗	88.3%	43.2	–
30	Qwen3-4B-Instruct-2507 thinkingAlibabaopen	82.5%	40.0	–
31	Qed NanoLM-Provers	82.5%	–	possible

More maths benchmarks · All LLM benchmarks

Cite as: BenchLeader, “AIME 2026 leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/matharena_aime_2026.

AIME 2026: questions

What does AIME 2026 measure?: The 2026 AIME I and II papers, evaluated by MathArena within days of the exam so that no model could have trained on them. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads AIME 2026?: GPT-5.5 leads AIME 2026 with 100.0% as of 25 Oct 2026, ahead of Claude Opus 4.8 at 100.0%.
How many models have AIME 2026 results?: 31 model configurations have a AIME 2026 result on BenchLeader, all taken from MathArena.
Who runs AIME 2026 and how often is it updated?: AIME 2026 is published by MathArena. BenchLeader re-reads the published results once a day and records the date each result was published.
Does AIME 2026 count toward the BenchLeader Index?: Yes. AIME 2026 contributes to the maths category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.