OpenAI AI Updates: May 22, 2026

1. AdventHealth deploys ChatGPT for Healthcare to offload administrative load from clinicians

OpenAI. OpenAI published an AdventHealth case study describing how the health system is rolling out ChatGPT for Healthcare to streamline administrative workflows and return clinician time to direct patient care. The post frames the deployment around AdventHealth’s whole-person care model, with ChatGPT used to draft documentation, summarize information, and reduce routine paperwork rather than make clinical decisions, sitting alongside the system’s existing electronic health record stack. It is one of the clearer signals that ChatGPT for Healthcare is moving from pilot territory into operational use at a large multi-state nonprofit hospital network with a meaningful compliance footprint. Source

2. Codex desktop ships Appshots, promotes Goal Mode to stable, and adds remote computer use from a locked Mac

OpenAI. The Codex desktop app rolled out an “Appshots” gesture where pressing both Command keys sends the frontmost app window to Codex as a screenshot plus extracted text, giving the agent immediate context on whatever the user is looking at without screen-share setup. Goal Mode graduated from experimental to stable across the macOS app, IDE extension, and CLI, letting Codex track long-running objectives that span multiple sessions, and a new Remote Computer Use capability lets Codex drive desktop applications while a Mac is locked, including remote access from Codex Mobile with safeguards. Teams on ChatGPT Business can now distribute reusable plugin bundles through marketplace sources, with Enterprise sharing on the way, and the in-app browser picked up faster image asset extraction plus better structured data handling on Windows and Chrome. Source

3. Codex CLI 0.133.0 turns Goals on by default and overhauls permission profiles, plugins, and remote control

OpenAI. Codex CLI 0.133.0 ships Goals enabled by default with dedicated storage and progress tracking, so multi-step objectives persist across runs without per-invocation setup, and codex remote-control now executes in the foreground with explicit machine status reporting. Permission profiles gain list APIs, inheritance, and managed requirements so administrators can compose policies instead of duplicating them, while plugin discovery becomes marketplace-aware with version visibility and extension lifecycle events expand to cover subagent start and stop plus async processing. Bug fixes resolve TUI working directory selection, plan-mode modified-Enter handling, stale background terminal events, raw code-mode output preservation, and AGENTS instruction loading with UTF-8 validation, alongside a canonical package archive pipeline, a Linux runtime wheel fix, and platform-sharded CI. Source

4. OpenAI reasoning model disproves 1946 Erdős conjecture, with Tao calling it ‘the first AI-generated argument I would not have spotted myself’

OpenAI. OpenAI said its frontier reasoning model produced a disproof of a long-standing Paul Erdős conjecture from 1946, with the result independently verified by working mathematicians and submitted to a top-tier journal. Fields Medalist Terence Tao called the proof “the first AI-generated argument I would not have spotted myself,” and the lab framed it as the first machine-led contribution to peer-reviewable mathematical research at a meaningful difficulty tier rather than an automated re-derivation of known results. The disclosure lands as OpenAI prepares its September IPO and continues to position reasoning-tuned models as the company’s frontier story, distinct from the consumer ChatGPT and Codex surfaces. Source