OpenAI has unveiled a research preview of Codex, a cloud-based AI coding agent designed to act as a virtual teammate for software developers.

OpenAI’s Latest Bet on the Future of Coding

OpenAI says Codex is its most advanced AI-powered software engineering agent to date. Codex has been designed to integrate directly into ChatGPT in order to assist with software development tasks ranging from writing features to fixing bugs. Available to Pro, Team, and Enterprise subscribers, and with support for Plus and Edu users expected in the near future.

AI Powered Tools Market

It’s worth noting here that Codex isn’t just another chatbot extension, but actually operates as a virtual, cloud-based coding assistant that can handle multiple tasks in parallel, provide verifiable output logs, and work with live codebases, all within a secure sandboxed environment. As such, it represents a clear move by OpenAI to secure a larger stake in the fast-expanding market for AI-powered developer tools.

So, What Is Codex, And What It’s For?

Codex is powered by codex-1, a specially trained variant of OpenAI’s o3 model, optimised for software engineering through reinforcement learning on real-world development tasks. OpenAI says this model is designed to produce cleaner, more human-readable code than its predecessors, and it can iteratively run and verify its own output using tests, linters, and other standard developer tools.

Once a user connects Codex to their GitHub repository, the agent loads the relevant files into a cloud sandbox and starts work. Users can assign coding tasks by typing a natural language prompt and clicking “Code”, or ask Codex questions about their codebase with the “Ask” function. Each job runs in an isolated environment, mimicking the structure and configuration of the user’s real-world dev setup.

Codex can:

– Write new features or functions

– Refactor and rename code

– Fix bugs and debug issues

– Answer questions about unfamiliar code

– Draft documentation and pull requests

– Run and verify tests.

Fast

Task completion using Codex typically takes between 1 and 30 minutes depending on complexity, and users can monitor progress and review results in real time. Once a job is done, Codex provides traceable logs of its actions and suggested changes, helping ensure transparency and accountability.

Safety

According to OpenAI Product Lead Alexander Embiricos, “a lot of the safety work from our o3 model carries over to Codex,” including the ability to reliably refuse malicious requests, such as writing malware, and restrict access to external APIs or the wider internet. This means Codex cannot be used to covertly access or manipulate live systems, though it also limits its utility for tasks requiring broader connectivity.

Who’s It For?

Right now, Codex is being rolled out to ChatGPT Pro, Enterprise, and Team subscribers worldwide. OpenAI says “generous access” will be granted initially, but usage will soon be capped by rate limits, with additional credits available for purchase. Plus and Edu users are next in line.

Tested Last Month

This preview follows months of behind-the-scenes work with early access testers. For example, companies such as Cisco, Temporal, Superhuman, and autonomous driving firm Kodiak have been helping OpenAI refine Codex in real-world settings. The use cases vary, but recurring themes include:

– Speeding up test coverage

– Automating background development tasks

– Enabling non-engineers to contribute lightweight code

– Improving developer focus and reducing context switching.

How Are The Testers Using It?

To give an idea of its real-world applications in some of these big-name tester companies, at Superhuman, Codex now helps product managers draft minor code changes without interrupting engineering teams. Also, at Kodiak (a company developing autonomous driving technology), it’s used to enhance test coverage and refactor key components in their autonomous driving stack. Also, at Temporal, which builds workflow tools for distributed applications, it supports debugging and iterative feature development.

A Follow-Up to Codex CLI

This latest launch builds on the release of Codex CLI last month, a lightweight, open-source command-line tool that lets developers interact with AI agents directly from their terminals. Codex CLI now defaults to codex-mini-latest, a variant of the o4-mini model optimised for low-latency editing and Q&A.

To streamline access, OpenAI has simplified authentication, i.e. users can now sign in via their ChatGPT accounts and instantly access API credits ($5 for Plus users, $50 for Pro) for a limited time. Pricing for codex-mini-latest is $1.50 per million input tokens and $6 per million output tokens, with a 75 per cent discount via prompt caching.

Broader Ambition

These updates appear to point to a broader vision, whereby OpenAI wants Codex to be more than just a helpful assistant, i.e. it wants Codex to function more like a trusted teammate. As Josh Tobin, OpenAI’s Agents Research Lead says – the goal is for Codex to eventually complete tasks autonomously that “take human engineers hours or even days”.

Riding the AI Coding Boom

This preview version is clearly a high-profile launch that OpenAI hopes will position it against growing rivals in the AI coding space, including Google, Microsoft, Anthropic, and a host of rapidly scaling startups. Also, more than that, it provides a glimpse into what the company sees as the future of human-AI collaboration in software engineering.

It should also be noted that the timing of Codex’s release is no accident. AI coding tools, sometimes dubbed “vibe coders”, have exploded in popularity in recent months. For example, both Google and Microsoft claim AI now contributes to around 30 per cent of their internal code output. Startups like Cursor have reached $300 million in annualised revenue and are being valued as high as $9 billion.

Meanwhile, OpenAI itself has reportedly just finalised a $3 billion deal to acquire Windsurf, the developer behind another major AI coding tool. Taken together, these moves appear to show the company is serious about expanding beyond its flagship chatbot and into the broader ecosystem of developer tools, including video generation (via Sora), research agents, and web automation.

Codex could, therefore, be a big part of that strategy. As AI capabilities advance, OpenAI seems to be envisioning a future where developers assign well-scoped tasks to multiple AI agents in parallel, review their work asynchronously, and move faster across every stage of the development lifecycle.

Challenges, Criticisms, and What Comes Next

Despite the buzz, Codex (like other AI coding agents) isn’t without its flaws. For example, a recent Microsoft study showed that even top-tier models such as Claude 3.7 Sonnet and o3-mini struggled to reliably debug software in complex environments. Codex’s own documentation notes that users must still manually review agent-generated code before integration or deployment.

Some developers have also expressed concern over workflow disruption. For example, delegating tasks to a remote agent adds latency compared to interactive editing, and current limitations, like the lack of image inputs for frontend work or the inability to mid-course correct an agent, may frustrate more advanced users.

There’s also the question of safety. While Codex runs in a secure, air-gapped environment and is trained to refuse malicious instructions, OpenAI acknowledges that balancing security with legitimate use cases (e.g. kernel-level programming) remains a work in progress.

Looking ahead, OpenAI says future versions of Codex will include more interactive features, deeper integration with developer tools (including CI systems and issue trackers), and even proactive collaboration capabilities, allowing agents to check in during a task and adapt based on new feedback. Also, if this research preview gains traction, it may change how businesses, from startups to enterprises, approach software development altogether.

What Does This Mean For Your Business?

Rather than offering a single-use tool or a generic assistant, OpenAI is positioning Codex as a foundational part of the software engineer’s toolkit, i.e. a collaborator that can work independently on defined tasks while staying closely aligned with human expectations and workflows. That positioning reflects OpenAI’s wider ambition to normalise multi-agent workflows and move towards asynchronous, AI-assisted development at scale.

For developers, especially those in high-pressure or fast-moving teams, Codex could help relieve the burden of repetitive or time-consuming tasks (refactoring code, writing tests, or updating documentation) freeing up more time for problem solving and creative development. At the same time, its ability to surface context, suggest improvements, and operate in secure containers could make it a valuable tool for navigating legacy codebases or tackling long-standing bugs. It may not be perfect, and as the documentation itself makes clear, human review is still essential, but the potential for improved focus and faster delivery could be really significant.

Businesses in particular may find this especially useful. For example, with ongoing skills shortages in the tech sector and continued demand for digital transformation, tools like Codex could help firms deliver projects faster without always needing to expand headcount. Small to medium-sized development teams, or organisations experimenting with agile methods, could see real benefits from delegating well-scoped tasks to Codex agents, either to accelerate delivery or free up senior engineers for more strategic work.

That said, the impact won’t be uniform. The very structure of Codex, i.e. running in isolated environments, without full real-time interactivity, means it will likely feel more useful for certain kinds of backend or infrastructure work than for frontend or UI-heavy tasks. Also, while Codex shows promising alignment with coding conventions and project practices, it’s still early days. Safety limitations, execution speed, and the need for well-prepared environments may restrict broader uptake until future iterations smooth out those rough edges.

For OpenAI, however, Codex is far more than a productivity tool. It’s a strategic move that places the company in direct competition with Microsoft’s GitHub Copilot, Google’s Gemini Code Assist, and Anthropic’s Claude Code. But unlike some of those offerings, Codex is being pitched not just as an assistant but as a long-term platform for agentic coding, one that could underpin an entire ecosystem of developer interactions, APIs, integrations, and workflows.

If early results hold up, Codex may well represent the next step in how code gets written, i.e. not just faster, but differently. By giving developers the tools to delegate, collaborate, and iterate with AI in the loop, OpenAI is betting on a future where productivity and creativity are amplified by intelligent software agents. And that’s a future which many UK businesses, whether they’re building the next app or maintaining critical infrastructure, may soon be part of.