2021 to 2026: Same Name, Two Entirely Distinct Products
On August 10, 2021, OpenAI launched Codex — a 12-billion-parameter model fine-tuned on 54 million public software repositories, capable of translating natural language instructions into runnable code. It served as the core engine behind GitHub Copilot and marked the first time countless developers were truly amazed by AI. Back then, Codex was essentially a code generator designed to help programmers write faster inline code completions.
However, in March 2023, this original Codex model was officially retired, replaced by more powerful general large language models GPT-3.5 and GPT-4. Surprisingly, even without dedicated code training, GPT-4 delivered superior code generation performance compared to the old Codex.
The real turning point arrived in 2025.
On May 16, 2025, OpenAI announced a brand-new iteration of Codex — no longer a standalone model, but a cloud-based software engineering agent. Codex entered a new agentic era.
The very first line of update notes on its official page read: “Update on May 16, 2025: We launched Codex, a cloud-based software engineering agent that can work on many tasks in parallel.”
This shift redefined Codex’s core mission from “completing code snippets for you” to “delivering full development projects end-to-end.” You only need to describe your goals in plain language — for instance, “Fix all lint errors in this repository” or “Implement a user login module.” Codex will autonomously write code, run unit tests and resolve bugs inside a cloud sandbox, finally submitting a complete Pull Request.
Internal OpenAI sources revealed that most of Codex’s own codebase was written by Codex itself. Alexander Embiricos, Codex Product Lead, told Ars Technica: “I think the vast majority of Codex is built by Codex, so it’s almost entirely just being used to improve itself.”
Three Product Offerings: CLI + Desktop App + IDE Plugins
The new Codex has built a full product ecosystem covering all development workflows:
Codex CLI (Open-Source Terminal Agent)
OpenAI open-sourced Codex CLI on April 16, 2025. It gained over 5,000 GitHub stars within 5 hours of release and hit 10,000 stars on its launch day. Greg Brockman, Co-Founder and President of OpenAI, stated at the time that Codex CLI was just the first of many upcoming open-source Agent tools. By 2026, Codex CLI had been fully rewritten from TypeScript to Rust, amassing more than 82,000 GitHub stars. It scored 77.3% on Terminal-Bench 2.0, a benchmark built specifically for native terminal AI coding agents, forming a three-way competitive landscape alongside Claude Code and Google’s CLI tool.
Codex Desktop App (Agent Command Center)
OpenAI rolled out a macOS-native desktop client in February 2026, officially branded the “Agent Command Center”. A Windows version followed in March the same year. The standout feature of this update is Codex’s ability to directly operate desktop applications on Mac: it controls an independent cursor to view screens, click UI elements and input text without human intervention. Developers leverage it for app testing and front-end iteration; multiple agents can run in parallel independently without disrupting your other work. Codex can also test web apps in browsers, retrieve context across tabs and utilize web DevTools, without taking full control of your browser window.
DeepIDE Plugin Ecosystem
Codex is integrated with mainstream IDEs including VS Code, Cursor and Windsurf via plugins, embedding autonomous coding capabilities directly into developers’ daily workspaces.
Full Integration Between ChatGPT and Codex On The Way
At the “Intelligence at Work” online conference on June 3, 2026, OpenAI officially announced that Codex would be merged into ChatGPT within the coming weeks. The two AI products with a combined user base of nearly one billion users will be fully unified.
Alexander Embiricos put it plainly: “You may not work 24/7, but your cloud-based agent will.”
This integration transforms Codex from an exclusive developer tool into an AI workforce accessible to everyone. It will be available across desktop, mobile and browsers, and embedded in familiar office software such as Excel and Slack as an on-demand all-purpose digital assistant. A preview version of Codex had already launched on the ChatGPT mobile app earlier, enabling asynchronous, on-the-go agentic coding workflows.
Simultaneously, OpenAI confirmed that all its cutting-edge models (GPT-5.5, GPT-5.4) and Codex now support one-click deployment on AWS Bedrock, allowing enterprise clients to access Codex’s full capabilities directly in the cloud.
Codex’s Three Flagship Upgrades: Agent Plugins + Annotations + Sites
OpenAI unveiled three core feature upgrades for Codex during the June 3 keynote:
Agent Plugins
OpenAI launched six role-specific enterprise plugins targeting core business verticals: data analytics, creative production, sales, product design, public market investment and investment banking. Each plugin bundles relevant tool integrations, domain expertise and operational workflows, connecting to 62 mainstream enterprise platforms (Salesforce, Snowflake, Figma, Canva, etc.) with a total of 110 built-in skills. Businesses can deploy AI specialists without any coding required. Weekly active Codex users surpassed 5 million in February this year, an 8x increase from the start of the year; knowledge workers adopted the tool three times faster than developers.
Annotations
Users can add direct annotations to all outputs generated by Codex, enabling targeted revisions with pinpoint accuracy and drastically lowering the barrier to modifying code. This functionality shines for cross-functional collaboration between non-technical stakeholders and AI agents.
Sites
With only a single URL input, Codex converts any document into an interactive website or app preview that can be instantly shared with teams and clients. Sites launched as a preview feature for enterprise and commercial teams on the same day, with a broader rollout to all subscription tiers planned later.
Technical Architecture: Cloud Sandboxes + Parallel Execution + Advanced Model Backbone
Codex is officially available on AWS Bedrock, powered by state-of-the-art models GPT-5.5, GPT-5.4 and the Codex software engineering agent, with flexible token-based pay-as-you-go pricing. GPT-5.5 excels at agentic coding, data analysis and multi-step autonomous task completion.
Its cloud sandbox infrastructure enables parallel execution of multiple isolated tasks. For five separate feature requests, Codex can process all five simultaneously in independent containers, whereas Claude Code requires manual orchestration of sub-agents to achieve parallelism. In terms of token efficiency, Codex consumes roughly one-third the tokens of Claude Code to complete equivalent tasks, delivering substantial cost advantages under API metered billing. Codex achieves approximately a 49% score on SWE-bench and a 90.2% pass rate on HumanEval.
Industry Landscape: Codex vs Claude Code
The AI coding agent market is currently dominated by two major competitors: OpenAI Codex and Anthropic Claude Code.
Core Advantages of Codex
Cloud-native architecture with isolated sandbox environments, purpose-built for task delegation and parallel workloads; superior performance on terminal automation tasks; unmatched token efficiency (3x better than Claude Code) for lower operational costs. Codex CLI is fully open-source, free to fork and customize for enterprise internal use cases.
Core Advantages of Claude Code
Leads Codex by roughly 23 percentage points on SWE-bench with a score of 72.5%, and hits a 92% pass rate on HumanEval. It delivers stronger performance on multi-file refactoring and large codebase overhauls, acting more like a senior engineer focused on robust code quality. Its massive 1-million-token context window outperforms Codex for long development sessions and extensive repository restructuring.
Usage Recommendations
Opt for Codex Desktop App or CLI for daily development and parallel task processing. Choose Claude Code when top-tier accuracy and complex large-scale code refactoring are priorities. ChatGPT Plus subscribers gain full access to Codex with zero additional fees.
Future Vision: Agentic Everything
During his appearance at the June 3 conference, Sam Altman predicted that the next major focus of the AI industry will be persistent, proactive autonomous agents running around the clock — his top priority area for investment over the coming year. He noted OpenAI is entering its third developmental phase: bringing AI access to every person.
From serving 3 million to 5 million developers and knowledge workers weekly, powering 100% of OpenAI’s internal codebase, and running an agentic code review system that processes over 100,000 external PRs daily, Codex is far more than a simple AI coding utility. It forms a complete agent ecosystem capable of managing the full software development lifecycle, shifting software creation from manual line-by-line coding to natural language task delegation handled autonomously in the cloud.
As the Codex product team shared on a podcast interview: “When building software becomes effortless, we will see an explosion of new applications. Custom-built tools tailored solely for you and your team will no longer be a luxury.”
Whether you are an independent developer looking to offload repetitive coding work, a technical lead evaluating the business costs of adopting AI coding agents, or a business team seeking to integrate Codex into corporate workflows, this cloud-native software engineering agent used by 5 million weekly users is well worth testing via ChatGPT Plus subscription, Codex Desktop App or open-source Codex CLI.