Evaluation sourcesOfficial evaluation
FrontierMath v2 · Tiers 1–3
Epoch AI / 推理 · 研究级数学推理
On HOTAIRanked
Evidence budget3.6%
Upstream data09/30 06:00
Last successful sync10/01 20:05
What and how it measures
只用 v2 的 Tiers 1–3；旧版题目和 Tier 4 不混入同一指标。
How this evidence is used
正式计分；同题家族共用固定预算。
Evaluation results
按固定规则，每个公开模型采用一套代表配置。匿名测试型号不展示，保留原榜名次，每项最多 30 个。
| Source rank | Source model | Raw score | Representative configuration |
|---|---|---|---|
| 1 | gpt-6.1-sol_maxOpenAI | 93.7% | Max 推理 |
| 1 | gpt-6-astra_maxOpenAI | 93.7% | Max 推理 |
| 3 | claude-opus-5-5_maxAnthropic | 91.2% | Max 推理 |
| 4 | claude-fable-5-1_maxAnthropic | 90.2% | Max 推理 |
| 5 | gpt-6-sol_maxOpenAI | 89.8% | Max 推理 |
| 6 | gpt-5.6-sol_maxOpenAI | 89.1% | Max 推理 |
| 7 | claude-sonnet-5-5_max— | 88.8% | Max 推理 |
| 8 | gpt-5.5-pro_xhighOpenAI | 87.7% | xHigh 推理 |
| 9 | claude-fable-5_maxAnthropic | 87.0% | Max 推理 |
| 10 | gpt-5.6-terra_maxOpenAI | 86.0% | Max 推理 |
| 11 | claude-opus-5_maxAnthropic | 85.6% | Max 推理 |
| 12 | gpt-5.5_xhighOpenAI | 85.3% | xHigh 推理 |
| 13 | gpt-5.4-pro-2026-03-05_xhighOpenAI | 82.5% | xHigh 推理 |
| 14 | gpt-5.6-luna_maxOpenAI | 82.1% | Max 推理 |
| 15 | claude-opus-4-8_maxAnthropic | 80.0% | Max 推理 |
| 16 | gpt-6-luna_maxOpenAI | 78.9% | Max 推理 |
| 17 | gpt-5.4-2026-03-05_xhighOpenAI | 78.6% | xHigh 推理 |
| 18 | qwen3.8-max_xhighAlibaba | 74.7% | xHigh 推理 |
| 20 | muse-spark-1.3_maxMeta | 74.0% | Max 推理 |
| 21 | gpt-5.2-pro-2025-12-11_xhighOpenAI | 74.0% | xHigh 推理 |
| 22 | kimi-k3_maxMoonshot AI | 72.2% | Max 推理 |
| 23 | gemini-3.7-flash_highGoogle | 71.6% | High 推理 |
| 24 | claude-opus-4-7_maxAnthropic | 70.2% | Max 推理 |
| 25 | glm-5.3_maxZ.ai | 68.8% | Max 推理 |
| 26 | gemini-3.8-flash_highGoogle | 68.4% | High 推理 |
| 27 | gpt-5.2-2025-12-11_xhighOpenAI | 67.4% | xHigh 推理 |
| 28 | grok-4.6_xhighxAI | 66.0% | xHigh 推理 |
| 28 | claude-opus-4-6_maxAnthropic | 66.0% | Max 推理 |
| 30 | qwen3.8-max-0902_xhighAlibaba | 65.6% | xHigh 推理 |
| 30 | claude-sonnet-5_maxAnthropic | 65.6% | Max 推理 |
Evaluation limits and attribution
正式计分；同题家族共用固定预算。
Data license: CC BY 4.0 · Epoch AI 自行评测数据
成绩由 Epoch AI 发布，原始分数与 HOTAI 共识分使用不同尺度，不能直接相加。