Taste-Bench benchmark shows GPT-5.6 Sol gets under 60% correct on long-term binary tasks
The Taste-Bench benchmark evaluates AI agents' strategic choices in long-running tasks. The top-performing model, GPT-5.6 Sol, selects the correct path on just under 60% of simple binary questions, and longer reasoning does not improve results.
Why it matters: It highlights current limitations of AI agents in long-term decision making, relevant for developers building autonomous systems.
