Все бенчмарки из новостей — от самых обсуждаемых к нишевым.
| # | Бенчмарк | Тип | Лидер | Моделей | Упом. |
|---|---|---|---|---|---|
| 01 | Terminal-Bench | agentic | 86.5% LLM-as-a-Verifier | 7 | 7 |
| 02 | GSM8K | math | 98% LLMs | 2 | 3 |
| 03 | OSWorld | agentic | 62.6% GPT-5.6 Sol | 3 | 3 |
| 04 | SWE-bench Verified | agentic | 81.27% SHERLOC | 3 | 3 |
| 05 | HarmBench | safety | 11.5% Opus 4.8 | 2 | 2 |
| 06 | MBPP+ | coding | 41% LLaDA-8B | 1 | 2 |
| 07 | PutnamBench | math | 587 Leanstral 1.5 | 1 | 2 |
| 08 | SWE-rebench | agentic | 56.5% Claude Opus 4.8 xhigh | 10 | 1 |
| 09 | ARC-AGI 2 | reasoning | 72.9% solver | 3 | 1 |
| 10 | JailbreakBench | safety | 93% STEER (applied to six open-source 8B-parameter models) | 2 | 1 |
| 11 | MATH-500 | math | 87% Qwen3.6-27B | 2 | 1 |
| 12 | τ²-bench | agentic | 0.355% ToolGraph+DPO | 2 | 1 |
| 13 | AIME 2024 | math | 62.4% Qwen3-1.7B | 1 | 1 |
| 14 | AIME 2025 | math | 50.8% DASH | 1 | 1 |
| 15 | BrowseComp | agentic | 92.2% GPT-5.6 Sol | 1 | 1 |
| 16 | SWE-bench | agentic | 80.3% Qwable-v1 | 1 | 1 |