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

Researchers have introduced a benchmark called Taste-Bench to test how well AI agents make strategic choices during long-running tasks. In results reported by The Decoder, the strongest performer, GPT-5.6 Sol, selects the right path on just under 60 percent of straightforward binary questions. Extending the models' thinking time does not close the gap.
The benchmark's focus is what might be called taste in decision-making: the ability to pick a sensible direction over the course of an extended task rather than to answer a single prompt correctly. Because the questions are simple binary choices, the sub-60-percent score suggests models struggle with the selection step itself, not with parsing complexity. The reported finding that longer reasoning fails to help points to a limitation that additional compute does not address, which is notable given how much current agent design relies on extended inference-time thinking.
For developers building agents, the takeaway is that raw reasoning capability may not translate into reliable judgment across long horizons. In practice, that argues for external checks, explicit decision criteria and human review at key branch points in agent workflows. This is my own reading of the implications rather than a reported conclusion.
Why it matters: Teams deploying agents for multi-step work should treat strategic choice as a separate risk from task execution, since even a leading model lands under 60 percent on simple binary decisions, according to the reported results. Benchmarking judgment separately from reasoning could become a standard part of agent evaluation.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
A new benchmark called Taste-Bench tests how well AI agents make strategic choices during long-running tasks. Even the best model, GPT-5.6 Sol, picks the right path on just under 60 percent of simple binary questions. Thinking longer doesn't help. The article AI agents have bad taste in long-term decisions and a new benchmark proves it appeared first on The Decoder .