Regex Golf
Generate the shortest valid regular expression matching all 10 target strings while excluding all 10 distractor strings across 25 deterministically generated puzzles. Lower is better.
Leaderboard
Intelligence Index- 15.3600
openai/gpt-6-astra
- 25.6400
google/gemini-3.8-flash
- 35.8000
openai/gpt-5.6-sol
- 46.0400
x-ai/grok-4.6
- 56.3600
google/gemini-3.7-flash
- 66.6400
anthropic/claude-opus-5
- 77.3200
anthropic/claude-fable-5
- 88.0000
openai/gpt-5.6-terra
- 98.4400
z-ai/glm-5.3-flash
- 1010.4400
openai/gpt-5.3-codex
- 1111.5600
openai/gpt-5.6-luna
- 1214.0400
thinkingmachines/inkling
- 1314.2800
z-ai/glm-5.3
- 1418.6800
anthropic/claude-sonnet-5
- 1530.7600
inclusionai/ling-3.0-flash
- 1631.4400
openai/gpt-oss-120b
- 1731.6800
anthropic/claude-opus-4.6
- 1839.7200
anthropic/claude-opus-4.5
- 1944.2000
anthropic/claude-opus-4.8
- 2044.7600
inception/mercury-2
- 2145.0000
anthropic/claude-opus-4.1
- 2252.6400
anthropic/claude-opus-4.7
- 2354.0400
moonshotai/kimi-k3
- 2454.5600
anthropic/claude-sonnet-4.5
- 2563.4800
anthropic/claude-haiku-4.5
- 2667.7600
openai/gpt-oss-20b
- 2770.6800
google/gemma-4-31b-it
- 2870.8400
openai/gpt-5.4
- 2973.1600
google/gemini-3-flash-preview
- 3079.8000
meta-llama/llama-4-maverick
- 3181.4800
openai/gpt-5.4-mini
- 3288.8400
anthropic/claude-sonnet-4.6
- 3390.8800
google/gemini-3.5-flash-lite
- 3492.9200
openai/gpt-5.1
- 3593.4800
deepseek/deepseek-v3.2
- 3694.2800
google/gemini-3.1-flash-lite-preview
- 37100.8800
x-ai/grok-4.20-beta
- 38104.7600
amazon/nova-pro-v1
- 39129.9600
mistralai/mistral-medium-3.1
- 40144.9200
mistralai/mistral-small-2603
- 41144.9600
upstage/solar-pro4
- 42153.9600
anthropic/claude-fable-5.1
- 43156.4000
amazon/nova-2-lite-v1
- 44168.2000
meta-llama/llama-4-scout
- 45217.6000
amazon/nova-micro-v1
- 46478.1200
amazon/nova-lite-v1
Prompt
The model receives the MATCH and REJECT string lists and must reply with only the raw regular expression in Python re syntax.
Score
Lower is better. Per puzzle: regex length plus 20 per missed MATCH, 20 per falsely matched REJECT, and a flat 200 for invalid regex or evaluation timeout. The benchmark score is the average across all 25 puzzles.
Execution
Benchmark runners execute locally, evaluate submissions with a sandboxed Python subprocess that is killed on timeout to catch catastrophic backtracking, use OpenRouter with reasoning output excluded, cache results in Neon, and skip recomputation for models that already have stored scores.