AI models show rapid gains on word and logic puzzles

AI models that solved only 18 percent of New York Times Connections puzzles in late 2024 now solve them nearly perfectly, signaling rapid progress in the technology's reasoning abilities.

By Middle East Affairs

September 2, 2026

An industrial concrete plant facility with blue and yellow equipment sits in a desert landscape with mountains in the background and solar panels visible to the right. — A concrete production facility with solar panels nearby, illustrating industrial infrastructure in an arid environment. (MIT Technology Review)

1 min read

Text size

Puzzles and games have long served as testing grounds for artificial intelligence development, from checkers in the 1950s to chess and Go in recent decades. Measuring progress on these tasks provides a window into both the strengths and limitations of AI systems.

Performance on word and logic puzzles has accelerated sharply in recent months. In late 2024, even the most advanced models could solve only 18 percent of New York Times Connections puzzles, but by early 2025, some models were solving them nearly perfectly every time.

The rapid gains reveal where AI systems excel and where human reasoning still outperforms them. Researchers have used a collection of seven puzzles to test where models succeed and fail against human solvers.