AI Models Advance to Complex Engineering and Automation Tasks

The Fable model set a new benchmark record by writing the most efficient GPU kernel tested on KernelBench-Mega, achieving an 18.71x speedup compared to an optimized PyTorch baseline. This achievement exemplifies how AI systems are autonomously developing fundamental components for AI research and development, potentially creating feedback loops of self-improvement. Progress in automation extends to real-world economic tasks. The Remote Labor Index, which evaluates AI systems' ability to execute online freelance projects end-to-end, showed dramatic improvement from 2.5% in October 2025 to 16.1% in July 2026. Recent models including Fable 5, GPT-5.5, and Opus 4.8 achieved success rates of 16.1%, 6.3%, and 8.3% respectively, indicating accelerating progress in economically relevant capabilities. Concurrently, researchers released OSWORLD 2.0, a benchmark assessing AI systems' competence in executing complex multi-program tasks spanning hours. The median task takes humans 1.6 hours, contrasting with 2 minutes in the previous version. Though current models achieve only 20.6% accuracy on complete tasks, performance improvement similar to the 1.0 benchmark is anticipated. In China, JD deployed the Oxygen AI Item Center to manage its vast inventory catalog, processing hundreds of millions of daily updates using Huawei Ascend NPUs, demonstrating how deep learning integrates with structured systems in production-scale applications.