SimpleBench

Where Everyday Human Reasoning Still Surpasses Frontier Models

SimpleBench Team

Introduction

We introduce SimpleBench, a multiple-choice text benchmark for LLMs where individuals with unspecialized (high school) knowledge outperform SOTA models. SimpleBench includes over 200 questions covering spatio-temporal reasoning, social intelligence, and what we call linguistic adversarial robustness (or trick questions). For the vast majority of text-based benchmarks LLMs outperform a non-specialized human, and increasingly, exceed expert human performance. However, on SimpleBench, a non-specialized human baseline is 83.7%, based on our small sample of nine participants, outperforming all 13 tested LLMs, including o1-preview, which scored 41.7%. While we expect model performance to improve over time, the results of SimpleBench confirm that the memorized knowledge, and approximate reasoning retrieval, utilized by frontier LLMs is not always enough to answer basic questions just yet.

Powered by Weave, from : Click here to learn more about Weave

Leaderboard

Rank Model Score (AVG@5) Organization
- Human Baseline* 83.7%
1st Claude 4 Opus (thinking) 58.8% Anthropic
2nd o3 (high) 53.1% OpenAI
3rd Gemini 2.5 Pro 51.6% Google
4th Claude 3.7 Sonnet (thinking) 46.4% Anthropic
5th Claude 4 Sonnet (thinking) 45.5% Anthropic
6th Claude 3.7 Sonnet 44.9% Anthropic
7th o1-preview 41.7% OpenAI
8th Claude 3.5 Sonnet 10-22 41.4% Anthropic
9th DeepSeek R1 05/28 40.8% DeepSeek
10th o1-2024-12-17 (high) 40.1% OpenAI
11th o4-mini (high) 38.7% OpenAI
12th o1-2024-12-17 (med) 36.7% OpenAI
13th Grok 3 36.1% xAI
14th GPT-4.5 34.5% OpenAI
15th Gemini-exp-1206 31.1% Google
16th Qwen3 235B-A22B 31.0% Alibaba
17th DeepSeek R1 30.9% DeepSeek
18th Gemini 2.0 Flash Thinking 30.7% Google
19th Llama 4 Maverick 27.7% Meta
20th Claude 3.5 Sonnet 06-20 27.5% Anthropic
21st DeepSeek V3 03-24 27.2% DeepSeek
22nd Gemini 1.5 Pro 002 27.1% Google
23rd GPT-4.1 27.0% OpenAI
24th GPT-4 Turbo 25.1% OpenAI
25th Claude 3 Opus 23.5% Anthropic
26th Llama 3.1 405b instruct 23.0% Meta
27th o3-mini (high) 22.8% OpenAI
28th Grok 2 22.7% xAI
29th Mistral Large v2 22.5% Mistral
30th Llama 3.3 70b instruct 19.9% Meta
31st DeepSeek V3 18.9% DeepSeek
32nd Gemini 2.0 Flash Exp 18.9% Google
33rd o1-mini 18.1% OpenAI
34th GPT-4o 08-06 17.8% OpenAI
35th Command R+ 17.4% Cohere
36th GPT-4o mini 10.7% OpenAI
temperature: 0.7, top-p: 0.95 (except o1 series)
*See Human Evaluation section of Report for details on how we calculated Human Baseline.
**We try an engineered prompt to optimize benchmark specific performance. See LLM Eval section of Report for details.

Video Summary

Evaluating Reasoning and Prompting

Performance comparison of different models on selected benchmarks

To assess LLMs fairly, we standardized prompts across all models, directing them to choose the most realistic answer step-by-step (COT). Additionally, we tested a benchmark specific engineered prompt for select models. Prompt engineering showed slight improvements suggesting that while tailored prompts can aid performance, fundamental limitations remain. In the full report, we also hypothesize that the surprising underperformance of GPT4o stems from optimizing for specific industrial applications (math and coding) at the expense of holistic reasoning.

For a deeper dive into our results and our methods, check out the full technical report here.