ThinkingBox benchmark: one-shot agent success rarely repeats

Microsoft authors released ThinkingBox-Bench: 507 business workflows across 5 domains, each run 20 times and graded on terminal backend state. Kimi-K3 solved 93.89% of tasks at least once but only 13.41% all 20 times, while Claude Opus 5 discovered fewer (79.09%) yet repeated far more (47.53%), showing discovery and repeatability rank models almost oppositely.

• Microsoft releases ThinkingBox to grade AI agents on database state2026-10-03
• AI Paradox Needs Guardrails, Not Panic2026-09-26
• Ternary Bonsai 2 (27B) released under 6GB2026-09-17
• Terminal Bench v4 ranks GLM-5.3 top2026-09-11
• Qwen3.8-27B Tuned for Humanlike Chat2026-09-11