1
anthropic/claude-opus-4.1
1.0000
WeirdBench
Benchmark leaderboard
ai-writing-detection
AI Writing Detection
Classify essays from a fixed balanced sample of 50 human-written and 50 AI-generated examples from the AI Generated Essays Dataset. Higher is better.
Higher score is better
2
anthropic/claude-opus-4.7
1.0000
3
google/gemini-3-flash-preview
1.0000
4
openai/gpt-5.5
1.0000
5
anthropic/claude-sonnet-4.5
0.9899
6
google/gemma-4-26b-a4b-it
0.9899
7
google/gemma-4-31b-it
0.9899
8
moonshotai/kimi-k2.5
0.9899
9
openai/gpt-5.3-chat
0.9899
10
qwen/qwen3.5-122b-a10b
0.9899
11
qwen/qwen3.5-397b-a17b
0.9899
12
openai/gpt-5.1
0.9804
13
x-ai/grok-4.1-fast
0.9804
14
anthropic/claude-opus-4.6
0.9800
15
anthropic/claude-sonnet-4.6
0.9592
16
anthropic/claude-opus-4.5
0.9278
17
moonshotai/kimi-k2.6
0.9009
18
google/gemini-3.1-flash-lite-preview
0.8929
19
inception/mercury-2
0.8750
20
openai/gpt-5.4
0.8621
21
anthropic/claude-haiku-4.5
0.8454
22
openai/gpt-oss-120b
0.8421
23
minimax/minimax-m2.5
0.8000
24
meta-llama/llama-4-maverick
0.7321
25
mistralai/mistral-small-2603
0.6849
26
openai/gpt-5.4-mini
0.6803
27
mistralai/mistral-large-2512
0.6667
28
x-ai/grok-4.20-beta
0.6667
29
amazon/nova-micro-v1
0.6438
30
mistralai/mistral-medium-3.1
0.6087
31
deepseek/deepseek-v3.2
0.4819
32
amazon/nova-2-lite-v1
0.4211
33
amazon/nova-pro-v1
0.0357
34
meta-llama/llama-4-scout
0.0241
35
amazon/nova-lite-v1
0.0000
Methodology
How scoring works
Read a fixed deterministic sample of 50 human essays and 50 AI essays from the downloaded Kaggle dataset, prompt the model once per essay to predict whether it is AI-generated, then compute binary classification precision, recall, and F1 for the AI class.
Prompt
Each essay is shown once and the model must return exactly one character: "1" for AI-generated or "0" for human-written, with no explanation.
Score
Higher is better. The benchmark score is the F1 score for detecting AI-generated essays, using label 1 as the positive class.
Execution
Benchmark runners execute locally, read the dataset from disk, use OpenRouter for predictions with reasoning disabled, cache results in Neon by benchmark and model ID, and skip recomputation for models that already have stored scores.