AI agents have bad taste in long-term decisions and a new benchmark proves it

A new evaluation tool called Taste-Bench reveals that current AI agents struggle to make effective strategic decisions during extended tasks. Even the top-performing model, GPT-5.6 Sol, succeeds on less than 60 percent of simple binary choices, regardless of how much time it spends processing. This benchmark highlights a persistent limitation in the ability of AI to maintain goal-oriented planning over the course of long-running operations.