Tops GPTs 
OpenAI ran Parameter Golf, an eight-week challenge with over 1,000 participants and 2,000+ submissions that asked teams to minimize held-out loss on the FineWeb dataset under a 16 MB artifact limit and a 10-minute training budget on 8×H100s. Organizers said AI coding agents widened participation but
Voyager has released a macOS file browser that runs AI agents inside folders and uses files as task context. The app supports Claude Code, Codex and Hermes; the maker says current features are free and file-change tracking is under research.
AutoScout24 Group says it cut development cycles from weeks to days by embedding Codex into engineering workflows and rolling out ChatGPT to roughly 2,000 employees. The company reported faster iterations, improved code quality and broader AI adoption among ~1,000 builders.
NVIDIA engineers report using Codex, built on GPT‑5.5 and running on NVIDIA GB200 and GB300, to speed development and automate research workflows. The company cites a 10x improvement in experiment turnaround and about 40k NVIDIANs with access to Codex.
OpenAI engineer David Wiesen explains a custom sandbox built to enforce file and network constraints for Codex on Windows. The design evolved from an unelevated prototype using SIDs and write-restricted tokens to an elevated model that adds firewall rules and dedicated local users.
OpenAI said on May 14, 2026 it updated ChatGPT to better detect evolving risk across conversations, targeting suicide, self-harm and harm-to-others. The changes add short safety summaries and model training; internal tests reported higher rates of safe responses.
OpenAI has added Codex to the ChatGPT mobile app in a preview, enabling live access to work running on laptops, devboxes, or remote environments. Remote SSH, Hooks, programmatic tokens and HIPAA support for eligible Enterprise workspaces are also detailed.
Sea Limited is deploying Codex across its engineering organisation, reporting 87% weekly active users. Co-Founder David Chen says AI agents are shifting developers’ roles and the company will host a regional Codex Hackathon Series with OpenAI across Asia.
Databricks reported that GPT‑5.5 set a new record on its OfficeQA Pro benchmark, achieving over 50% accuracy and reducing errors by 46% versus GPT‑5.4. The company is offering GPT‑5.5 through AI Unity Gateway for use in AgentBricks and the Agent Supervisor API.
Kuken Industries and Recursive built a field app using the OpenAI API that finds replacement-part drawings in about 4.5 minutes, down from nearly an hour. The company projects about a 90% reduction in hours spent searching drawings over a year.