OpenAI is merging ChatGPT, its coding agent Codex, and the developer API into a single product team led by Codex boss Thibault Sottiaux. The goal: a "super app" that also integrates the Atlas browser. Co-founder Greg Brockman officially takes over product strategy.

Oppo's Multi-X team released X-OmniClaw, an open-source agent that runs directly on Android devices and combines camera, screen, and voice to handle tasks in real apps. Instead of relying on cloud copies of the phone, the system uses local sensors; cloud compute only kicks in for reasoning. Tap paths get cloned as reusable skills, so the agent can jump straight to deeply nested app pages via deeplink next time around.

A consortium of 64 mathematicians built SOOHAK, a new AI benchmark with 439 handwritten tasks, including 99 that are deliberately unsolvable. Google's Gemini 3 Pro leads on research-level problems at 30 percent. But no model cracks 50 percent on spotting broken tasks. More compute makes models better at solving. It doesn't improve them at admitting a problem has no answer. SOOHAK tries to pin down the gap between a few flashy results and the broad research skills AI systems still lack.

World Action Models tackle a basic weakness of today's robotics AI: current models learn which movements match which camera images, but they don't understand how the world actually changes as a result. A new survey organizes about a hundred papers into two architectural lines and shows a key edge: these models can learn from everyday videos that contain no robot action labels. That kind of data was nearly useless for traditional robotics AI.