Benchmark leaderboard - 27 entries
AIME leaderboard
MathArena evaluates LLMs on the latest math competitions and olympiads; for AIME 2026 each model is run 4 times on each of the 30 final-answer problems and the average accuracy is reported. As of October 17, 2026, GPT-5.5 (xhigh) and Claude-Opus-4.8 (max) are tied for first on the AIME 2026 (MathArena) board with 100.00%.
Last verified from matharena.ai. Read the caveats
AIME 2026 (MathArena)
27 entries, in the board's order
| 1 | GPT-5.5 (xhigh) MathArena harness, 4 runs per problem, xhigh effort; rank 1 (tied); 95% CI ± 0.00%; warning icon shown on leaderboard; MathArena lists Final-Answer Comps as Deprecated Official leaderboard | 100.00% |
| 2 | Claude-Opus-4.8 (max) MathArena harness, 4 runs per problem, max effort; rank 1 (tied); 95% CI ± 0.00%; warning icon shown on leaderboard Official leaderboard | 100.00% |
| 3 | GPT-5.4 (xhigh) MathArena harness, 4 runs per problem, xhigh effort; rank 3; 95% CI ± 1.63%; warning icon shown on leaderboard Official leaderboard | 99.17% |
| 4 | GPT-5.2 (high) MathArena harness, 4 runs per problem, high effort; rank 4 (tied); 95% CI ± 2.29% Official leaderboard | 98.33% |
| 5 | Gemini 3.1 Pro Preview MathArena harness, 4 runs per problem; rank 4 (tied); 95% CI ± 2.29%; warning icon shown on leaderboard Official leaderboard | 98.33% |
| 6 | Step 3.5 Flash MathArena harness, 4 runs per problem; rank 6 (tied); 95% CI ± 3.21%; open weights, 196B total / 11B active Official leaderboard | 96.67% |
| 7 | Gemini 3 Flash MathArena harness, 4 runs per problem; rank 6 (tied); 95% CI ± 3.21% Official leaderboard | 96.67% |
| 8 | GLM 5 MathArena harness, 4 runs per problem; rank 6 (tied); 95% CI ± 3.21%; open weights, 744B total / 40B active; warning icon shown on leaderboard Official leaderboard | 96.67% |
| 9 | DeepSeek-v4-Pro (Max) MathArena harness, 4 runs per problem, Max mode; rank 6 (tied); 95% CI ± 3.21%; open weights, 1600B total / 49B active; warning icon shown on leaderboard Official leaderboard | 96.67% |
| 10 | Gemini 3.6 Flash MathArena harness, 4 runs per problem; rank 6 (tied); 95% CI ± 4.54%; warning icon shown on leaderboard Official leaderboard | 96.67% |
| 11 | Kimi K3 (Think) MathArena harness, 4 runs per problem, thinking mode; rank 6 (tied); 95% CI ± 4.54%; open weights, 2800B total / 50B active; warning icon shown on leaderboard Official leaderboard | 96.67% |
| 12 | Claude-Opus-4.6 (High) MathArena harness, 4 runs per problem, High effort; rank 6 (tied); 95% CI ± 3.21% Official leaderboard | 96.67% |
| 13 | DeepSeek-v4-Flash (Max) MathArena harness, 4 runs per problem, Max mode; rank 13 (tied); 95% CI ± 3.58%; open weights, 248B total / 13B active; warning icon shown on leaderboard Official leaderboard | 95.83% |
| 14 | Kimi K2.5 (Think) MathArena harness, 4 runs per problem, thinking mode; rank 13 (tied); 95% CI ± 3.58%; open weights, 1000B total / 32B active Official leaderboard | 95.83% |
| 15 | GLM 5.1 MathArena harness, 4 runs per problem; rank 13 (tied); 95% CI ± 3.58%; open weights, 744B total / 40B active; warning icon shown on leaderboard Official leaderboard | 95.83% |
| 16 | Kimi K2.6 (Think) MathArena harness, 4 runs per problem, thinking mode; rank 13 (tied); 95% CI ± 3.58%; open weights, 1000B total / 32B active; warning icon shown on leaderboard Official leaderboard | 95.83% |
| 17 | Claude-Opus-4.7 (xhigh) MathArena harness, 4 runs per problem, xhigh effort; rank 13 (tied); 95% CI ± 3.58%; warning icon shown on leaderboard Official leaderboard | 95.83% |
| 18 | Step 3.7 Flash MathArena harness, 4 runs per problem; rank 18 (tied); 95% CI ± 3.90%; open weights, 198B total / 11B active; warning icon shown on leaderboard Official leaderboard | 95.00% |
| 19 | Gemini 3.5 Flash MathArena harness, 4 runs per problem; rank 18 (tied); 95% CI ± 5.51%; warning icon shown on leaderboard Official leaderboard | 95.00% |
| 20 | Grok 4.1 Fast (Reasoning) MathArena harness, 4 runs per problem, reasoning mode; rank 20 (tied); 95% CI ± 4.19% Official leaderboard | 94.17% |
| 21 | DeepSeek-v3.2 (Think) MathArena harness, 4 runs per problem, thinking mode; rank 20 (tied); 95% CI ± 4.19%; open weights, 671B total / 37B active Official leaderboard | 94.17% |
| 22 | Qwen3.5-397b-a17b MathArena harness, 4 runs per problem; rank 20 (tied); 95% CI ± 4.19%; open weights, 397B total / 17B active; warning icon shown on leaderboard Official leaderboard | 94.17% |
| 23 | Qwen3.5-35B-A3B MathArena harness, 4 runs per problem; rank 23; 95% CI ± 4.46%; open weights, 35B total / 3B active; warning icon shown on leaderboard Official leaderboard | 93.33% |
| 24 | Qwen3.5-9B MathArena harness, 4 runs per problem; rank 24; 95% CI ± 4.71%; open weights, 9B; warning icon shown on leaderboard Official leaderboard | 92.50% |
| 25 | NVIDIA-Nemotron-3-Super MathArena harness, 4 runs per problem; rank 25 (tied with Qwen3.5-27B and Gemini 3 Pro (preview)); 95% CI ± 4.95%; open weights, 120B total / 12B active; warning icon shown on leaderboard Official leaderboard | 91.67% |
| 26 | Qwen3.5-27B MathArena harness, 4 runs per problem; rank 25 (tied); 95% CI ± 4.95%; open weights, 27B; warning icon shown on leaderboard Official leaderboard | 91.67% |
| 27 | Gemini 3 Pro (preview) MathArena harness, 4 runs per problem; rank 25 (tied); 95% CI ± 4.95% Official leaderboard | 91.67% |
About this board
- AIME 2025 leaderboard (https://matharena.ai/?comp=aime--aime_2025) not recorded in this pass; only AIME 2026 was read.
- MathArena lists all Final-Answer Comps (including AIME 2025 and AIME 2026) as Deprecated, so current frontier models (GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol, Gemini 4 Argon, Claude-Opus-5.5, Grok 4.7, Qwen3.8) have no AIME 2026 entry on the official leaderboard.
- Per-model evaluation dates are not shown on the AIME 2026 leaderboard; date column is the read date 2026-10-17.
- Rank 25 is a three-way tie, so 27 entries are returned instead of 25.
- Six entries below rank 25 (GLM 5.2 90.00%, Qwen3.5-4B 89.69% with IRT-estimated scores, Qwen3-30B-A3B-2507-Think 88.33%, QED-Nano 82.50%, Qwen3-4B-2507-Think 82.50%) were left out.
- Vendor-reported AIME numbers from model cards were not collected because the official leaderboard was available.
FAQ
What does AIME measure?
MathArena evaluates LLMs on the latest math competitions and olympiads; for AIME 2026 each model is run 4 times on each of the 30 final-answer problems and the average accuracy is reported.
Which model leads AIME?
As of October 17, 2026, GPT-5.5 (xhigh) and Claude-Opus-4.8 (max) are tied for first on the AIME 2026 (MathArena) board with 100.00%. That is the board's own order, not a Developers Digest ranking.
Where do these AIME scores come from?
All 27 entries come from the official leaderboard at matharena.ai. Every entry links to its source, which was read on October 17, 2026, and lists the setting it was run under.
Related reading
- VibeThinker-3B: A 3 Billion Parameter Model That Outscores Opus 4.5 on ReasoningVibeThinker-3B is a 3B parameter reasoning model reporting 94.3% on AIME 2026. What the paper claims, how to run it locally, and what it trades away.
- Kolibri Release Guide: Aleph Alpha's 78B Open ModelKolibri is an Apache 2.0 English-German MoE with 78B total and 3B active parameters and a 1M context. Benchmarks, the vLLM command, and who should run it.
- Muse Glimmer 30B: Meta's Open-Weight Local Agent Model, Benchmarks, and Hardware RealityMuse Glimmer 30B is Meta's Apache 2.0 open-weight multimodal agent model that runs in 24GB. Benchmarks, hardware reality and what the numbers actually say.
- Inkling: Thinking Machines Lab Drops a 975B Open-Weights ModelA new American open-weights frontier model with multimodal capabilities, 1M token context, and competitive benchmarks. Here's what the HN community thinks.
- GLM-5.2 Local Deployment: Running Z.ai's 744B Model on Consumer HardwareRun GLM-5.2 locally with Unsloth dynamic quantization: hardware math for a 256GB Mac or 24GB GPU, quantization tradeoffs and lessons from the HN thread.
- Mellum2 Developer Guide: JetBrains' Open-Source Coding ModelMellum2 developer guide: JetBrains' open-source 12B MoE coding model with 2.5B active parameters. How to run it locally and where it fits in your coding stack.
Sources
Last verified . Each row links to the page its score was copied from.
Newsletter
Get new benchmark results as they land
One email per week. New model releases, the leaderboard moves that matter and what they mean for what you build.