FrontierMath v2 Tiers 1-3 (FrontierMath v2 (Tiers 1-3))
Epoch AI's corrected v2 core FrontierMath suite of private advanced mathematics problems. Models can reason iteratively and use Python; scores are pass rates on the private set.
Top models on FrontierMath v2 (Tiers 1-3) — October 29, 2026
As of October 29, 2026, GPT-6 Astra leads the FrontierMath v2 (Tiers 1-3) leaderboard with 93.684% , followed by GPT-6.1 Sol (93.684%) and Claude Opus 5.5 (91.228%).
GPT-6 Astra
OpenAI
GPT-6.1 Sol
OpenAI
Claude Opus 5.5
Anthropic
68 modelsMathematics30% of category scoreCurrentUpdated October 29, 2026
Leaderboard (68 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAI | 93.684% | Not reported | Closed |
| 2 | GPT-6.1 SolOpenAI | 93.684% | Not reported | Closed |
| 3 | Claude Opus 5.5Anthropic | 91.228% | Not reported | Closed |
| 4 | GPT-6.1 TerraOpenAI | 90.877% | Not reported | Closed |
| 5 | Claude Fable 5.1Anthropic | 90.175% | Not reported | Closed |
| 6 | GPT-6 SolOpenAI | 89.825% | Not reported | Closed |
| 7 | GPT-5.6 SolOpenAI | 89.000% | Not reported | Closed |
| 8 | Claude Sonnet 5.5Anthropic | 88.772% | Not reported | Closed |
| 9 | GPT-5.5 ProOpenAI | 87.719% | Not reported | Closed |
| 10 | Claude Fable 5Anthropic | 87.018% | Not reported | Closed |
| 11 | Gemini 4 ArgonGoogle | 85.965% | Not reported | Closed |
| 12 | Claude Opus 5Anthropic | 85.614% | Not reported | Closed |
| 13 | GPT-5.5OpenAI | 85.263% | Not reported | Closed |
| 14 | GPT-5.6 TerraOpenAI | 84.900% | Not reported | Closed |
| 15 | GPT-5.4 ProOpenAI | 82.456% | Not reported | Closed |
| 16 | Claude Opus 4.8Anthropic | 80.000% | Not reported | Closed |
| 17 | GPT-6 LunaOpenAI | 78.947% | Not reported | Closed |
| 18 | GPT-5.6 LunaOpenAI | 78.600% | Not reported | Closed |
| 19 | GPT-5.4OpenAI | 78.596% | Not reported | Closed |
| 20 | Claude Haiku 5.5Anthropic | 75.088% | Not reported | Closed |
| 21 | Qwen3.8 MaxAlibaba | 74.737% | Not reported | Open |
| 22 | Muse Spark 1.3Meta | 74.035% | Not reported | Closed |
| 23 | Gemini 3.7 FlashGoogle | 71.579% | Not reported | Closed |
| 24 | Claude Opus 4.7Anthropic | 70.175% | Not reported | Closed |
| 25 | GLM-5.3Z.AI | 68.772% | Not reported | Open |
| 26 | Gemini 3.8 FlashGoogle | 68.421% | Not reported | Closed |
| 27 | GPT-5.2OpenAI | 67.400% | Not reported | Closed |
| 28 | DeepSeek V4.1 Flash 1017DeepSeek | 67.368% | Not reported | Open |
| 29 | Claude Opus 4.6Anthropic | 65.965% | Not reported | Closed |
| 30 | Grok 4.6xAI | 65.965% | Not reported | Closed |
| 31 | Claude Sonnet 5Anthropic | 65.614% | Not reported | Closed |
| 32 | Qwen3.7 MaxAlibaba | 64.561% | Not reported | Closed |
| 33 | Gemini 3.5 FlashGoogle | 62.807% | Not reported | Closed |
| 34 | Gemini 3.1 ProGoogle | 59.649% | Not reported | Closed |
| 35 | GLM-5.2Z.AI | 59.206% | Not reported | Open |
| 36 | Gemini 3.6 FlashGoogle | 58.947% | Not reported | Closed |
| 37 | Kimi K2.6Moonshot AI | 57.193% | Not reported | Open |
| 38 | Grok 4.5xAI | 57.193% | Not reported | Closed |
| 39 | GLM-5.3-FlashZ.AI | 55.789% | Not reported | Open |
| 40 | Kimi K2.7 CodeMoonshot AI | 54.035% | Not reported | Open |
| 41 | Grok 4.7xAI | 52.982% | Not reported | Closed |
| 42 | GPT-5.4 miniOpenAI | 51.228% | Not reported | Closed |
| 43 | Gemini 3 FlashGoogle | 51.228% | Not reported | Closed |
| 44 | GPT-5 miniOpenAI | 46.667% | Not reported | Closed |
| 45 | Inkling-SmallThinking Machines Lab | 46.316% | Not reported | Open |
| 46 | DeepSeek V4 ProDeepSeek | 45.263% | Not reported | Open |
| 47 | GPT-5.4 nanoOpenAI | 44.912% | Not reported | Closed |
| 48 | Grok 4.3xAI | 42.807% | Not reported | Closed |
| 49 | Qwen3.6 PlusAlibaba | 38.246% | Not reported | Closed |
| 50 | GLM-5.1Z.AI | 36.842% | Not reported | Open |
| 51 | o4-mini (high)OpenAI | 36.140% | Not reported | Closed |
| 52 | Qwen3.6-27BAlibaba | 35.088% | Not reported | Open |
| 53 | Qwen3.7 PlusAlibaba | 34.386% | Not reported | Closed |
| 54 | Claude Opus 4.5Anthropic | 34.386% | Not reported | Closed |
| 55 | InklingThinking Machines Lab | 33.333% | Not reported | Open |
| 56 | o3OpenAI | 33.333% | Not reported | Closed |
| 57 | Gemini 3.1 Flash-LiteGoogle | 27.719% | Not reported | Closed |
| 58 | Gemini 3.5 Flash-LiteGoogle | 25.965% | Not reported | Closed |
| 59 | Gemini 2.5 ProGoogle | 24.561% | Not reported | Closed |
| 60 | Claude Sonnet 4.5Anthropic | 23.860% | Not reported | Closed |
| 61 | Qwen3.6-35B-A3BAlibaba | 20.351% | Not reported | Open |
| 62 | GPT-5 nanoOpenAI | 20.000% | Not reported | Closed |
| 63 | Qwen3.7 FlashAlibaba | 19.298% | Not reported | Closed |
| 64 | Qwen3.5 FlashAlibaba | 18.246% | Not reported | Closed |
| 65 | o1OpenAI | 14.737% | Not reported | Closed |
| 66 | GPT-4.1 miniOpenAI | 6.667% | Not reported | Closed |
| 67 | GPT-4.1OpenAI | 5.965% | Not reported | Closed |
| 68 | GPT-4oOpenAI | 0.351% | Not reported | Closed |
According to BenchLM.ai, GPT-6 Astra leads the FrontierMath v2 (Tiers 1-3) benchmark with a score of 93.684%, followed by GPT-6.1 Sol (93.684%) and Claude Opus 5.5 (91.228%). The top models are clustered within 2.456 points, suggesting this benchmark is nearing saturation for frontier models.
68 models have been evaluated on FrontierMath v2 (Tiers 1-3). The benchmark falls in the Mathematics category. The Mathematics leaderboard ranks models by a weighted category score, and FrontierMath v2 (Tiers 1-3) contributes 30% of it. BenchAlign v5.8 also uses it in the overall ranking.
About FrontierMath v2 (Tiers 1-3)
Year
2026
Tasks
285 private advanced mathematics problems
Format
Python-enabled iterative mathematical problem solving
Difficulty
From olympiad-plus to early research mathematics
After Epoch AI's June 2026 correction, FrontierMath v2 Tiers 1-3 contains 295 problems, including ten public examples. Published scores use the 285 private problems. We select the highest published reasoning-effort result for each model and keep this core score distinct from Tier 4.
Freshness and provenance
Version
FrontierMath v2 (Tiers 1-3) 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does FrontierMath v2 (Tiers 1-3) measure?
Epoch AI's corrected v2 core FrontierMath suite of private advanced mathematics problems. Models can reason iteratively and use Python; scores are pass rates on the private set.
Which model scores highest on FrontierMath v2 (Tiers 1-3)?
GPT-6 Astra by OpenAI currently leads with a score of 93.684% on FrontierMath v2 (Tiers 1-3).
How many models are evaluated on FrontierMath v2 (Tiers 1-3)?
68 AI models have published results on FrontierMath v2 (Tiers 1-3) in the BenchLM catalog.
Compare top models on FrontierMath v2 (Tiers 1-3)
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.