Who leads the MATH benchmark?

o4-mini (high) currently leads MATH with a score of 98.2 on accuracy.

What is the state-of-the-art score on MATH?

The state-of-the-art result on MATH is 98.2 (accuracy), achieved by o4-mini (high) as of 2026.

How many models are tracked on MATH?

Codesota tracks 50 models on MATH.

When was the MATH leaderboard last updated?

The MATH leaderboard on Codesota includes results through 2026, with the earliest tracked result from 2023.

Codesota · Benchmark · MATHHome/Leaderboards/Language & Knowledge/Mathematical Reasoning/MATH

Unknown

MATH.

Name: MATH Benchmark Results
Creator: Unknown
Published: 2023-01-01
License: https://creativecommons.org/licenses/by/4.0/

12,500 competition mathematics problems (5,000 test) from AMC, AIME, and other sources covering algebra, geometry, number theory, and more. Harder than GSM8K. Modern evaluations typically use the MATH-500 representative subset.

Paper ↗Leaderboard ↓Lineage

§ 01 · SOTA history

Year over year.

§ 02 · Leaderboard

Results by metric.

Found a wrong score or missing run?

Use row edits to send a sourced correction into moderation.

Add / edit result ↗Report issue ↗

accuracy

Accuracy is the reported evaluation metric for MATH. Codesota tracks published model scores on this metric so readers can compare state-of-the-art results across sources and model families.

Higher is better

Trust tiers for accuracyverifiedpapervendorcommunityunverified

Rank	Model	Trust	Score	Year	Links	Fix
01	o4-mini (high) MATH-500, zero-shot CoT, pass@1. High reasoning effort.	paper	98.2	2026	Source ↗	Looks wrong?
02	o3 (high) MATH-500, zero-shot CoT, pass@1. High reasoning effort.	unverified	98.1	2026	Source ↗	Looks wrong?
03	o3-mini MATH-500, zero-shot CoT, pass@1. High reasoning effort.	paper	97.9	2026	Source ↗	Looks wrong?
04	o3 MATH-500, zero-shot CoT, pass@1. Default reasoning effort.	unverified	97.8	2026	Source ↗	Looks wrong?
05	o4-mini MATH-500, zero-shot CoT, pass@1. Default reasoning effort.	unverified	97.5	2026	Source ↗	Looks wrong?
06	Gemini 2.5 Pro MATH-500, pass@1. Gemini 2.5 Pro (Mar 2025).	paper	97.3	2026	Source ↗	Looks wrong?
07	DeepSeek R1 MATH-500, pass@1. From official DeepSeek-R1 paper (Jan 2025).	paper	97.3	2026	Source ↗	Looks wrong?
08	DeepSeek-R1 MATH-500, pass@1. From official DeepSeek-R1 paper (Jan 2025).	paper	97.3	2026	Source ↗	Looks wrong?
09	o1 MATH-500, zero-shot CoT, pass@1.	paper	96.4	2026	Source ↗	Looks wrong?
10	Kimi k1.5 MATH-500, long-CoT variant. From official Kimi k1.5 paper (Jan 2025).	paper	96.2	2026	Source ↗	Looks wrong?
11	Claude 3.7 Sonnet MATH-500 with extended thinking enabled.	paper	96.2	2026	Source ↗	Looks wrong?
12	DeepSeek-R1-Zero MATH-500, pass@1. DeepSeek-R1-Zero (pure RL, no SFT). From R1 paper (Jan 2025).	paper	95.9	2026	Source ↗	Looks wrong?
13	DeepSeek-R1-Distill-Llama-70B MATH-500, pass@1. Distilled from DeepSeek-R1 into Llama-3.1-70B. From R1 paper (Jan 2025).	paper	94.5	2026	Source ↗	Looks wrong?
14	DeepSeek-R1-Distill-Qwen-32B MATH-500, pass@1. Distilled from DeepSeek-R1 into Qwen-2.5-32B. From R1 paper (Jan 2025).	paper	94.3	2026	Source ↗	Looks wrong?
15	DeepSeek-v3-0324 MATH-500. DeepSeek-V3-0324 updated model (Mar 2025). Non-reasoning base model.	unverified	94	2026	Source ↗	Looks wrong?
16	Claude Opus 4.5 4-shot. Source: Claude Opus 4.5 model card, Anthropic (2025).	verified	90.7	2026	Source ↗	Looks wrong?
17	QwQ-32B MATH-500, pass@1. QwQ-32B reasoning model by Alibaba/Qwen (Mar 2025).	unverified	90.6	2026	Source ↗	Looks wrong?
18	DeepSeek-V3 MATH-500. Non-reasoning base model. From DeepSeek-V3 technical report (Dec 2024).	paper	90.2	2026	Source ↗	Looks wrong?
19	o1-mini MATH-500, zero-shot CoT, pass@1.	paper	90	2026	Source ↗	Looks wrong?
20	Llama 4 Maverick 4-shot. Source: Meta Llama 4 model card (April 2025).	verified	89.4	2026	Source ↗	Looks wrong?
21	Claude Opus 4 4-shot. Source: Claude Opus 4 model card, Anthropic (2025).	verified	89.2	2026	Source ↗	Looks wrong?
22	Claude Sonnet 4 4-shot. Source: Claude Sonnet 4 model card, Anthropic (2025).	verified	88.9	2026	Source ↗	Looks wrong?
23	GPT-4.5 Preview Full MATH test set, zero-shot CoT.	paper	87.1	2026	Source ↗	Looks wrong?
24	o1-preview MATH-500, zero-shot CoT, pass@1.	paper	85.5	2026	Source ↗	Looks wrong?
25	Qwen2.5-Plus	unverified	84.7	2024	Paper ↗Code ↗	Looks wrong?
26	Qwen2.5-72B-Instruct Qwen2.5-72B-Instruct. Table 6 in Qwen2.5 Technical Report.	verified	83.1	2026	Source ↗	Looks wrong?
27	Qwen2.5-VL-72B	unverified	83	2025	Paper ↗Code ↗	Looks wrong?
28	GPT-4.1 Full MATH test set, zero-shot CoT.	paper	82.1	2026	Source ↗	Looks wrong?
29	MiniMax-Text-01	unverified	77.4	2025	Paper ↗Code ↗	Looks wrong?
30	gpt-4o Full MATH test set, zero-shot CoT. gpt-4o-2024-05-13.	paper	76.6	2026	Source ↗	Looks wrong?
31	Grok 2 Full MATH test set.	paper	76.1	2026	Source ↗	Looks wrong?
32	Llama 3 (405B, Instruct)	unverified	73.8	2024	Paper ↗Code ↗	Looks wrong?
33	Llama 3.1 405B Full MATH test set.	paper	73.8	2026	Source ↗	Looks wrong?
34	GPT-4 Turbo Full MATH test set, zero-shot CoT.	paper	73.4	2026	Source ↗	Looks wrong?
35	Qwen3-235B-A22B	unverified	71.84	2025	Paper ↗Code ↗	Looks wrong?
36	claude-35-sonnet Full MATH test set. Original Claude 3.5 Sonnet (June 2024).	paper	71.1	2026	Source ↗	Looks wrong?
37	Claude 3.5 Sonnet Full MATH test set. Original Claude 3.5 Sonnet (June 2024).	unverified	71.1	2026	Source ↗	Looks wrong?
38	gpt-4o-mini Full MATH test set, zero-shot CoT.	paper	70.2	2026	Source ↗	Looks wrong?
39	GPT-4o mini Full MATH test set, zero-shot CoT.	unverified	70.2	2026	Source ↗	Looks wrong?
40	Llama 3.1 70B Full MATH test set.	unverified	68	2026	Source ↗	Looks wrong?
41	gemini-15-pro From Google's official evaluation.	paper	67.7	2026	Source ↗	Looks wrong?
42	Gemini 1.5 Pro From Google's official evaluation.	unverified	67.7	2026	Source ↗	Looks wrong?
43	Step-3.5-Flash Base	unverified	66.8	2026	Paper ↗Code ↗	Looks wrong?
44	Claude 3 Opus Full MATH test set.	unverified	60.1	2026	Source ↗	Looks wrong?
45	HRM-Text-1B	unverified	56.5	2026	Paper ↗Code ↗	Looks wrong?
46	Aria	unverified	50.8	2024	Paper ↗Code ↗	Looks wrong?
47	Apertus-70B-Instruct	unverified	30.8	2025	Paper ↗Code ↗	Looks wrong?
48	Chameleon 34B	unverified	22.5	2024	Paper ↗Code ↗	Looks wrong?
49	LLaMA-65B	unverified	20.5	2023	Paper ↗Code ↗	Looks wrong?
50	SmoLM2 (1.7B)	unverified	11.6	2025	Paper ↗Code ↗	Looks wrong?

Lineage

MATH in context.

See full mathematical reasoning benchmarks lineage →

Predecessors (1)

saturated2021-10

GSM8K

GSM8K addressed grade-school arithmetic; MATH jumped directly to AMC/AIME/Olympiad competition problems — a 5-difficulty-level span above GSM8K. Released one month apart, they together defined the 2021–2023 math evaluation landscape.

This benchmark (1)

saturating2021-11

MATH

Successors (2)

active2024-03

AIME 2024

AIME is not an AI benchmark by origin — it is the human competition that feeds into USAMO/IMO. It became an AI frontier benchmark when o1 started scoring competitively. Because it is updated annually with fresh problems, it provides contamination control that MATH (fixed dataset) cannot.

active2024-10

OmniMATH

OmniMATH aggregates olympiad-level problems from IMO and national competitions, providing harder instances than MATH's Level 5 problems. A parallel ceiling-raising branch alongside AIME.

§ 04 · Submit a result

Add to the leaderboard.

← Back to Mathematical Reasoning