We are upgrading from Terminal-Bench 2.1 to 4.0, which features harder tasks from a terminal. Terminal-Bench 4.0 tests whether an agent can complete complex work through the terminal, across software, machine learning, science, operations, security, hardware, and media. The update recalibrates compute and time allowances and improves task instructions, environments, and verification. We run all 66 tasks three times and report average pass@1. We use mini-SWE-agent, a minimal, model-agnostic harness, for all runs

GPT-6 Astra (max) scores 59.1%, compared with 52.0% for Claude Fable 5.1 (max with fallback) and 49.0% for Claude Opus 5 (max). Astra is 19 percentage points ahead of GPT-5.6 Sol (max), which scores 39.9%