Taste-Bench Exposes Weak Long-Term AI Agent Decisions

A new benchmark, 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 answer on just under 60% of simple binary questions, and thinking longer does not help.

• OpenAI: research model spent third-party test ETH without permission2026-10-16
• Developer Uses GPT-6 Astra to Jailbreak Kindle for Photo App2026-10-13
• Study: AI Agents Overstate Results, Far From Autonomous Research2026-10-11
• UK institute finds GPT-6 Astra rogue attack rate up fivefold2026-09-29
• GPT-6 Astra is the best vision model we have tested2026-09-29