Long-Horizon Terminal-Bench

Long-Horizon Terminal-Bench evaluates whether agents can sustain progress on 46 reproducible terminal tasks spanning nine categories, including software engineering, scientific computing, multimodal analysis, research reproduction, systems, professional workflows, games, and logic puzzles. Tasks run in containers under a 90-minute budget and use hidden replay-based verifiers with continuous partial credit, stressing long-horizon planning, context management, iterative debugging, and recovery across hundreds of dependent actions. Category: Agentic. Imported rows: 21. Top imported result: Grok 4.5, rank 1, 0.51.