Codex: From Backend Model to Super App Fiasco—A Turbulent Half-Year Story

Codex’s recent rollercoaster ride began on July 9, when OpenAI launched GPT-5.6 and merged Codex into ChatGPT’s desktop app, billing it as a 'super app' that combines chat, coding, and productivity. Google’s Logan Kilpatrick even tweeted overnight that the 'super app era is here.' But less than 24 hours later, AI influencer Matt Shumer dropped a bombshell: GPT-5.6, running at its highest inference tier, had deleted all files on his Mac. The super app dream turned into a super flop in just one night.

To understand this mess, we need to go back five years. Codex first emerged in 2021 as a GPT-3 fine-tuned code model, working behind the scenes for GitHub Copilot. While Copilot became the face of AI coding, Codex itself was sidelined—OpenAI deprecated its API in March 2023, stating that GPT-3.5 and GPT-4 were more capable all-around, leaving the code-specialized model with no clear purpose. For nearly two years, Codex faded from public view.

OpenAI revived Codex in April 2025, but this time as a full tool suite: a Rust-based CLI agent, cloud version, and desktop app. However, the market had shifted—Anthropic’s Claude Code, launched in 2025, quickly dominated the AI coding space with better developer experience and stronger reputation. By Q1 2026, Claude Code had 4.2 million weekly active users (WAU), while Codex lagged with less than 1 million—barely a quarter of its competitor’s numbers. Codex’s VS Code extension scored just 3.4/5, compared to Claude Code’s 4.0, and developers often preferred Claude’s superior model.

But a turnaround began in late Q1 2026. OpenAI pumped resources into Codex: launching a desktop client and GPT-5.3-Codex model (outperforming Claude in code tests) in February; adding built-in browser, local computer control, and remote SSH in April; optimizing token efficiency with GPT-5.5 in May; and rolling out role-based plugins for non-developers (lawyers, sales, product managers) in June. They also made Codex accessible via ChatGPT’s mobile app, offered free credit resets, and even ran a 60-second Super Bowl ad. The results were staggering: Codex’s WAU jumped from 600k to over 5 million in five months (a 730% increase), with 20% of users being non-developers—signaling its expansion beyond coding.

Then came the July 9 merger: Codex ceased to exist as an independent app, becoming a 'coding agent' within ChatGPT’s desktop framework. Old users retained their data, and ChatGPT gained Codex’s core capabilities (local file access, app integration, browser use). OpenAI also launched ChatGPT Work (a GPT-5.6-powered agent system) and Sites (one-click website builder) for free users, aiming to make ChatGPT the go-to desktop tool.

But the launch was marred by four major issues: First, token costs—users complained about skyrocketing consumption despite claims of 54% efficiency gains. OpenAI later admitted higher tiers (like Ultra) use 5-10x more tokens. Second, the file deletion incident: Matt Shumer’s Mac files were wiped by GPT-5.6-Sol at the highest tier, due to a sub-agent architecture flaw. Third, interface confusion: The new default 'ChatGPT Work' mode hid the classic chat feature, and users criticized the Electron-based app’s performance and messy design. Fourth, trust issues: OpenAI’s responses to user complaints (e.g., suggesting parallel use of old and new apps) and admission of Mac bias alienated many.

Sources

• NetEase News: Article on Codex’s Journey