1
openai/gpt-5.3-codex
3.6000
WeirdBench
Benchmark leaderboard
wordle
Wordle
Play 20 recent Wordle answers turn by turn with standard gray/yellow/green feedback. Invalid guesses still cost a turn, scores are capped at 10 turns per puzzle, and lower is better.
Lower score is better
2
openai/gpt-5.3-chat
3.8000
3
openai/gpt-5.5
4.0000
4
anthropic/claude-opus-4.6
4.0500
5
openai/gpt-oss-120b
4.2000
6
anthropic/claude-opus-4.7
4.3000
7
inception/mercury-2
4.5500
8
anthropic/claude-sonnet-4.5
4.6500
9
anthropic/claude-opus-4.5
5.5500
10
google/gemini-3-flash-preview
6.8000
11
anthropic/claude-opus-4.1
6.8500
12
google/gemini-3.1-flash-lite-preview
8.2000
13
moonshotai/kimi-k2.6
8.7000
14
anthropic/claude-haiku-4.5
9.2000
15
openai/gpt-5.4
9.3500
16
google/gemma-4-31b-it
9.6000
17
x-ai/grok-4.20-beta
9.6000
18
deepseek/deepseek-v3.2
9.7000
19
openai/gpt-5.4-mini
9.7000
20
mistralai/mistral-medium-3.1
9.7500
21
openai/gpt-5.1
9.7500
22
mistralai/mistral-large-2512
9.8500
23
google/gemma-4-26b-a4b-it
9.9500
24
meta-llama/llama-4-maverick
9.9500
25
amazon/nova-2-lite-v1
10.0000
26
amazon/nova-lite-v1
10.0000
27
amazon/nova-micro-v1
10.0000
28
amazon/nova-pro-v1
10.0000
29
google/gemini-3.1-pro-preview
10.0000
30
meta-llama/llama-4-scout
10.0000
31
minimax/minimax-m2.5
10.0000
32
minimax/minimax-m2.7
10.0000
33
mistralai/mistral-small-2603
10.0000
34
moonshotai/kimi-k2.5
10.0000
35
openai/gpt-oss-20b
10.0000
36
qwen/qwen3.5-27b
10.0000
37
stepfun/step-3.5-flash:free
10.0000
38
xiaomi/mimo-v2-pro
10.0000
39
z-ai/glm-5
10.0000
Methodology
How scoring works
Use a fixed set of 20 recent Wordle answers, run a fresh chat loop for each puzzle, and score the average turns needed to solve while applying standard duplicate-letter feedback rules.
Prompt
The model is told to reply with exactly one 5-letter word per turn, that any extra text is penalized, and that duplicate letters are allowed.
Score
Lower is better. Each puzzle score is the turn the word is solved on, or 10 if the model never solves it within 10 turns. Invalid guesses still count as turns.
Execution
Benchmark runners execute locally, simulate the Wordle judge deterministically, cache results in Neon, and skip recomputation for models that already have stored scores.