ThinkingBox puts agent reliability at 25.25%

Microsoft’s ThinkingBox benchmark found GPT-5.4 completed all 20 repeated attempts on just 25.25% of 507 stateful business tasks, even though it found a successful trajectory at least once on 91.12%, according to Microsoft’s benchmark report. The gap says retries can discover a path without making an agent dependable; builders should measure repeatable state changes and recovery, not just a green final response, extending the archive’s agent-reliability analysis.