# MathArena Apex 2025 ![Category](https://img.shields.io/badge/Category-Math-1f6feb) ![Metric](https://img.shields.io/badge/Metric-Accuracy-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-medium-d29922) ![Created](https://img.shields.io/badge/Created-2025-555) [![Paper](https://img.shields.io/badge/Paper-arXiv-b31b1b)](https://arxiv.org/abs/2505.23281) [![GitHub](https://img.shields.io/badge/GitHub-eth--sri%2Fmatharena-181717?logo=github&logoColor=white)](https://github.com/eth-sri/matharena) [![Leaderboard](https://img.shields.io/badge/Leaderboard-MathArena-0969da)](https://matharena.ai/) > **Quick answer:** MathArena Apex 2025 is a competition-mathematics configuration within ETH SRI's MathArena platform — final-answer problems from the "Apex" 2025 competition set, evaluated using MathArena's standard anti-contamination protocol of averaging results over four independent runs. ## At a Glance **What it tests:** Final-answer competition mathematics, drawn from a curated 2025 competition ("Apex") set within the broader MathArena platform. **Why it matters:** MathArena's whole premise is evaluating models on math competitions that are unlikely to already be in training data, giving a cleaner read on reasoning ability than older, widely-leaked competition benchmarks. **Known limitations:** No authoritative public item count for this specific configuration has been published; treat it as one dated configuration within MathArena's continuously evolving set, not a fixed standalone benchmark. ## What MathArena Apex 2025 Measures MathArena is a continuously updated platform built by researchers at ETH Zurich's SRI lab and INSAIT to evaluate LLMs on math competitions that resist training-data contamination. The "Apex 2025" configuration is one specific competition set tracked within that platform — final-answer competition mathematics questions, graded automatically and averaged over repeated runs (MathArena's standard protocol uses four runs per problem) to reduce single-sample variance. A high score indicates strong competition-level mathematical problem-solving on a set specifically chosen to minimize the chance the model has memorized the answers from pretraining data. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Math | | Metric | Final-answer accuracy (avg. of 4 runs) | | Number of tasks | Not publicly documented as a fixed count for this configuration | | Saturation | Medium | | Created by | MathArena / ETH SRI | | Source paper | [MathArena team 2025/2026](https://arxiv.org/abs/2505.23281) | | GitHub | [eth-sri/matharena](https://github.com/eth-sri/matharena) | | Leaderboard | [matharena.ai](https://matharena.ai/) | ## How MathArena Apex 2025 Is Scored Each problem has an automatically-checkable final answer. MathArena's standard protocol runs each model four times per problem and reports the average accuracy, reducing variance from any single sampling run. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [Hy4 Preview](/models/hungyuan/hy4-preview) | 74.2% | [Tencent Hunyuan model card](https://huggingface.co/tencent/Hy4-preview) | 2026-08 | *Scores sourced from published technical reports and model cards. Results depend on harness, prompt format, and effort settings — see each source for methodology.* ## MathArena Apex 2025 on Benchgen > No Benchgen results yet — [be the first to run MathArena Apex 2025](/benchmarks/matharena-eth-sri/matharena-apex-2025). ## MathArena Apex 2025 vs Other Benchmarks | Benchmark | What it tests | Tasks | Saturation | |---|---|---|---| | **MathArena Apex 2025** | Apex 2025 competition mathematics | — | Medium | | [ArXivMath](/benchmarks/matharena-eth-sri/arxivmath) | Fresh math from new arXiv papers | Rolling | Low | | [HMMT 2026](/benchmarks/hmmt-2026) | Static competition math | — | Medium | | [AIME 2026](/benchmarks/aime-2026) | Static competition math (AIME) | — | High | Use MathArena Apex 2025 alongside ArXivMath for a fuller picture of MathArena's anti-contamination approach to math evaluation — competition-style problems here, freshly-sourced paper-derived problems there. ## Run MathArena Apex 2025 on Your Model Benchgen lets teams run MathArena Apex 2025 against their own model versions, compare results across runs, and catch regressions in competition-math reasoning — rather than relying on a single vendor-reported number. ## Related Models & Benchmarks - Top performers: [Hy4 Preview](/models/hungyuan/hy4-preview) - Related benchmarks: [ArXivMath](/benchmarks/matharena-eth-sri/arxivmath) · [BrokenArXiv](/benchmarks/matharena-eth-sri/brokenarxiv) ## Frequently Asked Questions

What is MathArena Apex 2025?

MathArena Apex 2025 is a competition-mathematics configuration within ETH SRI's MathArena platform, evaluating final-answer problems from the Apex 2025 competition set.

What does a good MathArena Apex 2025 score look like?

As of August 2026, frontier models score in the 66–91% range; scores above 74% represent strong competition-math performance.

Who created MathArena?

MathArena was created by researchers at ETH Zurich's SRI lab and INSAIT, including Dekoninck, Jovanović, Gehrunger, Rögnvaldsson, Petrov, Sun, and Vechev.

MathArena Apex 2025 — Results

MathArena Apex 2025 — Evaluations