engblogs

A detailed audit reveals pervasive evaluation flaws in SWE-Bench Pro and related benchmarks, arguing for stronger signal, longer-horizon tasks, improved data quality checks, human-in-the-loop review, and cautious deployment decisions under OpenAI’s Preparedness Framework.

GPT-Live is a continuous, full-duplex voice system that talks and listens in real time, delegates complex tasks to frontier models like GPT-5.5, and integrates provenance signals and watermarking with a verification API to enable safer, more natural ChatGPT Voice experiences.

OpenAI articulates National Security Principles to guide government partnerships and frontier AI use in national security, emphasizing transparency, democratic accountability, meaningful human judgment, and safeguards across cyber defense and biosecurity contexts.

OpenAI Academy's AI Skills Jam for K-12 educators is a hands-on, in-person program with mentor support that helps teachers and district leaders translate AI into practical, scalable improvements in teaching, planning, and communication.

A rigorous audit of SWE-Bench Pro and its predecessor reveals widespread evaluation flaws and demonstrates how to separate signal from noise in coding benchmarks, outlining a QA-driven pipeline that leverages both automated and human reviews to ensure fair, meaningful model capability assessments.

OpenAI's National Security Principles outline responsible government and national security partnerships for frontier AI, emphasizing transparency, democratic accountability, and safeguards while enabling cyber defense and biosecurity applications.

Grok 4.5 delivers the strongest mean performance on ~2,000 GDPval+ professional tasks, outperforming GPT 5.5 and Opus 4.8 across domains with lower error rates and expert-rubric satisfaction.

Grok 4.5 delivers the strongest mean pass rate on Snorkel’s GDPval+ professional-work tasks, outperforming GPT 5.5 and Opus 4.8 across domains like legal, education, healthcare, and QA while showing lower error rates and more robust failure handling.

A case study on automating cross-repo documentation with GitHub Agentic Workflows, mapping product milestones to release branches and delivering SME-reviewed docs via secure, token-scoped pull requests in a separate docs repo.

Automates cross-repo, SME-reviewed documentation by generating PRs from feature changes via GitHub Agentic Workflows, using a constrained per-workflow bot and a safe-outputs pipeline to securely publish docs across repos.

MUFG partners with OpenAI to roll out ChatGPT Enterprise across Mitsubishi UFJ Bank, accelerating an AI-native transformation with enterprise-wide adoption and AI-powered customer experiences in a secure, governed digital bank future.

AP+ accelerates product development, investigations, and member communications by deploying ChatGPT Enterprise and Codex to speed complex issue resolution, simulate payment journeys, and strengthen governance in a regulated payments ecosystem.