MathArena Apex 2025 — Results
| Rank | Model | Score |
|---|---|---|
| 1 | hy4-preview | 74.2 |
MathArena Apex 2025
1 phaseActive
MathArena's Apex 2025 competition-math configuration — final-answer scoring on ETH SRI's continuously updated, anti-contamination math platform.
Overview
MathArena Apex 2025
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 |
| GitHub | eth-sri/matharena |
| Leaderboard | 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 | 74.2% | Tencent Hunyuan model card | 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.
MathArena Apex 2025 vs Other Benchmarks
| Benchmark | What it tests | Tasks | Saturation |
|---|---|---|---|
| MathArena Apex 2025 | Apex 2025 competition mathematics | — | Medium |
| ArXivMath | Fresh math from new arXiv papers | Rolling | Low |
| HMMT 2026 | Static competition math | — | Medium |
| 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
- Related benchmarks: ArXivMath · BrokenArXiv