Making Connections: Artificial Intelligence meets Abstract Ingenuity using NYT Connections by Uyi Uwadiae
What happens when you pit eight leading AI models against The New York Times Connections puzzle?
In this data-driven experiment, I tested free versions of Gemini, Claude, DeepSeek, Vibe by Mistral, Copilot, ChatGPT, Muse, and Grok across 10 rigorous rounds of Connections—combining 5 historically difficult published NYT games with 5 original custom puzzles designed to test zero-shot lateral thinking.
Key Highlights & Findings:
The Top Tier: Gemini and Claude tied for the most connections found (30/40), with Claude taking the top spot when factoring in partial credit for "one-away" guesses.
The Custom Puzzle Wall: While models successfully cleared 91% of official NYT boards, they hit a 100% impenetrable cognitive wall on non-semantic, visual, and structural custom puzzles (0/8 on themed boards) due to tokenization blind spots.
Model Personalities & Quirks: From Gemini’s aggressive "Purple-first" hunting to ChatGPT's 16 near-miss "one-aways" and Copilot's hilarious Windows version hallucinations, each model exhibited a distinct behavioral play style.
Ultimately, this benchmark illustrates why generative AI excels at semantic pattern retrieval, but why true lateral creativity, visual intuition, and abstract wordplay remain uniquely human.
Check out my companion essays and the Crafting Connections series on Substack and Medium!