PlatinumBench (MIT): leaderboard
Carefully curated benchmarks testing LLM reliability - manually revised to remove dataset errors. Measures consistency and correctness across math, logic, reading comprehension, and commonsense tasks.
Metric: Avg Error Rate (%). Source: platinum-bench.csail.mit.edu. Status: saturated. 34 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Pro (Preview) | 0.53 |
| 2 | O1 (High) | 0.74 |
| 3 | Grok 3 | 0.79 |
| 4 | GPT-4.5 (Preview) | 0.99 |
| 5 | Claude 3.7 Sonnet | 1.05 |
| 6 | Claude 3.5 Sonnet | 1.09 |
| 7 | GPT-4.1 (2025-04-14) | 1.11 |
| 8 | O1 (Medium) | 1.27 |
| 9 | DeepSeek R1 | 1.4 |
| 10 | Grok 3 Mini (High) | 1.51 |
| 11 | Claude 3.5 Sonnet (20240620) | 1.84 |
| 12 | Llama 3.1 405B Instruct | 1.92 |
| 13 | Qwen 2.5 Max | 2.05 |
| 14 | Llama 4 Maverick | 2.12 |
| 15 | GPT-4o (2024-08-06) | 2.24 |
Interactive version: theaggregate.ai/benchmark?slug=platinumbench-mit · How It Works · Data refreshed daily, snapshot 2026-09-05.