02:09
2026-08-26
byteiota.com
ai-research
Microsoft ThinkingBox Exposes AI Agent Reliability Gap
Microsoft released ThinkingBox, an open-source benchmark and sandbox, on August 19, 2026, revealing that the best AI models, such as GPT-5.4, achieve only 25.25% reliability across 20 attempts on busi…