AI · Settled
Will OpenAI have the best Code Arena | WebDev AI at the end of September 2026?
The answer: NO
The panel of 20 models averaged 43% YES; 14 of 20 were on the right side. The crowd price when the panel was sealed: 65%.
THE BRIEF, WRITTEN AFTER THE OUTCOME
OpenAI did not win Code Arena WebDev AI; panel mean 43% vs Polymarket 65% YES
The outcome was NO. The panel averaged 43% YES, while the Polymarket price frozen at question time was 65% YES. The best panel call was DeepSeek V3 + web search at 17%, which correctly favoured NO but recorded no thesis or drivers. The biggest misses were the high-YES models, for example GPT-5 + web search at 80%, none of which recorded supporting arguments in the records.
- Outcome: NO, resolved 2026-09-30 18:15:01.
- Panel mean_p_yes 0.43, median_p_yes 0.4; 14 of 20 models called it correctly.
- Crowd was a Polymarket price frozen at 0.65 YES.
- Best single model Brier among correct calls: DeepSeek V3 + web search, 17% YES, Brier 0.029.
- Largest wrong call: GPT-5 + web search, 80% YES, Brier 0.64.
EVERY FORECAST, HIGHEST FIRST
- GPT-5 + web search80%
- Perplexity Sonar Pro72%
- DeepSeek R170%
- Perplexity Sonar69%
- Mistral Large62%
- GPT-5 Mini57%
- GPT-546%
- Claude Sonnet 4.5 + web search42%
- Llama 4 Maverick42%
- Kimi K242%
- Gemini 2.5 Pro38%
- Qwen3 235B32%
- Claude Opus 4.132%
- DeepSeek V330%
- Grok 4.528%
- Claude Sonnet 4.528%
- Gemini 2.5 Flash25%
- Gemini 2.5 Pro + web search24%
- Llama 4 Maverick + web search20%
- DeepSeek V3 + web search17%