OpenAI has released a research preview of a powerful new AI agent designed to streamline software engineering tasks, marking a significant step forward in one of AI’s fastest-growing application areas.

The tool, called Codex, can handle multiple development responsibilities simultaneously—including writing code, debugging, running tests, and answering questions about a company’s codebase. Built on a specialized version of OpenAI’s o3 reasoning model—dubbed codex-1—this agent is tailored specifically for software engineering workflows. It’s available to ChatGPT Pro, Team, and Enterprise users.

With this release, OpenAI is signaling a stronger commitment to AI-driven coding—an area that’s rapidly gaining traction across the tech industry. Major players like Microsoft, Amazon, Google, Anthropic, and rising startups like Anysphere (creator of the developer-favorite Cursor) have all been ramping up investments in this space.

“There’s a lot of great AI out there,” said Srinivas Narayanan, OpenAI’s VP of Engineering. “We’re definitely seeing intense competition.”

Adding to the momentum, OpenAI is reportedly in acquisition talks with AI coding startup Windsurf, with a potential valuation of $3 billion, according to Bloomberg. Neither OpenAI nor Windsurf commented on the report.

Despite OpenAI’s dominant position in consumer AI tools, it’s still building its reputation among developers. According to Wayne Chi and Valerie Chen, PhD researchers at Carnegie Mellon University who study coding model performance, OpenAI’s status in the coding world lags behind competitors like Anthropic, whose Sonnet models are currently favored by many developers.

“Sonnet has some real staying power,” said Chi. “But the landscape is evolving so quickly, it’s hard to crown a definitive winner just yet.”

OpenAI noted that its newer GPT-4.1 model has been rapidly adopted and has topped performance benchmarks in certain coding scenarios—though not all involve deep reasoning tasks.

Codex, the company said, will redefine how developers work by automating time-consuming processes. Unlike most real-time coding assistants, Codex runs independently in the cloud and returns output within 1 to 30 minutes. OpenAI also claims the tool excels at learning and adapting to an organization’s unique coding style, making it especially useful for code reviews and style consistency.

“This represents a fundamentally different way of working,” said Narayanan. He emphasized that the research preview is a phased rollout meant to gather user feedback, minimize risks, and refine the technology before a wider release.

“We’re still in the early days,” he added. “There’s a lot we need to learn.”

Source