OpenAI moves toward 'work agents' and controlled access

OpenAI is remaking Codex from a coding assistant into a software agent, while putting its new biology model behind a vetted-access gate. (openai.com) On April 16, OpenAI said the updated Codex app for macOS and Windows can control desktop apps, browse inside the app, generate images, remember user preferences, and handle ongoing or repeatable work. The company also said Codex now supports reviewing pull requests, viewing multiple files and terminals, and connecting to remote development boxes over Secure Shell. (openai.com) Two days earlier, OpenAI expanded its Agents software kit, and in March it released GPT-5.4 with built-in computer use, tool search, and a 1 million-token context window for longer workflows. The product line now points in one direction: models that act across tools instead of waiting for one prompt at a time. (openai.com ) (openai.com) OpenAI made the same move on the life-sciences side. On April 16 it introduced GPT-Rosalind as a research preview in ChatGPT, Codex, and the application programming interface for “qualified customers” under what it calls a trusted access program. (openai.com) The company said GPT-Rosalind is designed for biological reasoning, scientific tool use, and multi-step work such as target discovery, genomics interpretation, pathway analysis, literature synthesis, and hypothesis generation. OpenAI also launched a Codex plugin that connects researchers to more than 50 scientific tools and data sources. (openai.com) (help.openai.com) OpenAI paired that release with restrictions. It said trusted access customers are vetted, deployments include biosecurity controls, and the company is starting with a limited research preview rather than a broad public rollout. (openai.com) That is a sharper commercial split than the ChatGPT era made visible: general-purpose models are getting wider distribution, while higher-risk domain systems are arriving through enterprise channels, plugins, and access controls. OpenAI’s April 8 strategy post also said its next phase of enterprise AI centers on products for work, teams, and specialized business use. (openai.com 1) (openai.com 2) The timing overlaps with management churn. Bloomberg reported on April 3 that Chief Operating Officer Brad Lightcap moved into a special projects role and two senior executives took medical leave, and CNBC reported on April 18 that three more OpenAI executives announced departures. (bloomberg.com) (cnbc.com) OpenAI has not publicly announced a GPT-5.5 launch date on its newsroom or product pages as of April 18. That leaves outside reports about timing as speculation, even as the company keeps shipping the pieces around it: stronger coding models, desktop agents, and controlled scientific access. (openai.com 1) (openai.com 2) The picture from this week is less about one model name than about how OpenAI wants its software used: not just to answer questions, but to do work inside guarded lanes. (openai.com 1) (openai.com 2)