How Codex Swallowed ChatGPT: Inside OpenAI's 10-Month Agent Turnaround

Background: GPT-5 Preview and the Claude Code Contrast

In summer 2025, Every co-founder Dan Shipper tested GPT-5 inside Codex and found it fast and smart but overly cautious — stopping early like a pair-programming partner. By contrast, Every's team had already switched to Claude Code (Opus 4), handing off entire features and walking away. Shipper confronted Codex product lead Alexander Embiricos: GPT-5 seemed to be building a perfect tool for a dying workflow.

The Two Threads: Feeling AGI and Bootstrapping

Codex's revival followed two internal threads. Embiricos, after joining via Multi's acquisition, first "felt AGI" when a reasoning model was hooked into a terminal — reading files, running commands, fixing errors autonomously. Thibault Sottiaux (Tibo), a research-infrastructure veteran from DeepMind and OpenAI, started a separate effort in December 2024: training an internal coding agent on OpenAI's monorepo to accelerate infrastructure work. Researchers used it daily, creating a tight feedback loop. Tibo described this as bootstrapping : researchers use the model to run experiments, producing better models that further improve the agent. When asked about splitting a separate model line, Tibo said no such decision existed — it began as a research exploration for speed.

Internal Startup Model and Parallel Bets

By late 2024, coding became a company-wide OKR. Greg Brockman noted the unresolved question was product form factor: terminal, editor sidebar, or cloud VM. Sam Altman framed OpenAI as betting on a portfolio of internal startups. Multiple teams pursued different forms in parallel — a necessity in fast-moving AI, since models capable of 30-second reasoning suit sidebars, while hour-long autonomy demands a new shape. Teams eventually merged, combining product thinkers (Embiricos), engineering leads (Tibo), and researchers (Mia Glaese).

Cloud Launch Failure (May 2025)

After Anthropic released Claude Code (terminal-based) in February 2025, Codex bet on the cloud. In ~7 weeks they shipped Codex Cloud (May 16): describe a task in ChatGPT, Codex works in an isolated cloud VM, returns a PR button. Internally it thrived on OpenAI's standardized environment; externally it flopped. Developers spent hours replicating local configs; tasks couldn't be interrupted mid-flight; each follow-up spun a new VM. Embiricos admitted the model wasn't smart enough; Glaese reflected they built for a future vision without meeting users on their laptops. Tibo revealed they had a terminal prototype but abandoned it after a month to go all-in on cloud — the ideal was a cloud AI with its own laptop. Meanwhile, an engineering manager left for Cursor, and the rumored $3B Windsurf acquisition fell through.

Pivot to Local: Rust Rewrite and Model Improvements

Tibo took over and redirected to local-first . Instead of piling features onto a CLI, he rewrote the agent core in Rust , decoupling file/command primitives from the terminal UI so the same agent could run in editors, desktop apps, cloud, or ChatGPT. On the model side, Michelle Pokrass's team made GPT-5 better at real-world coding by shadowing engineers and turning failure points into training tasks. This explained the "timidity" Shipper observed: burned by cloud, OpenAI built a cautious collaborator — exactly what most engineers wanted. Post-GPT-5, VS Code and Cursor extensions saw 10x+ usage growth by October. Embiricos declared the rebuilt local agent finally had a model that could drive it.

Desktop App Without an Editor

Defying convention, the team invested heavily in a native desktop app with no built-in code editor . Tibo saw long-running, multi-agent sessions making 1970s-style terminals absurd. Embiricos envisioned a workflow of close collaboration then gradual delegation. Designer Andrew Ambrosino drew inspiration from chat apps, not VS Code, and made the bold call: no manual editing — users must delegate all changes to the agent. The bet: win by making the agent so good developers stop missing an editor. Early days were painful: the team built the app in VS Code while dogfooding it; daily "are you using it as your main tool?" checks yielded universal "no." Complaints included writing prompts for trivial edits. When Google launched Antigravity (with a full editor) in November, the team questioned their fork-less approach. Model quality was also a blocker — Ambrosino felt the November model couldn't support a release regardless of UI.

Internal Adoption and Merger with ChatGPT

Turnarounds arrived rapidly: GPT-5.1 Codex Max (Nov 19) coded continuously for hours; GPT-5.2 Codex (Dec 18) excelled in large repos. Because the Codex model line was independent, the team iterated fast without risking 800M ChatGPT users. By January the whole team switched; by February researcher Noam Brown "fully switched," and Brockman (terminal/Emacs veteran) called returning to old workflows "like going back in time." Altman later called it OpenAI's most internally loved product ever. On Feb 2, 2026, the macOS desktop app launched — multiple agents working on independent code copies. Shipper noted features weren't novel (Claude Desktop had them) but Codex felt cleaner, faster, and truly agent-native. GPT-5.3 Codex (days later) expanded beyond coding to general knowledge work. Shipper himself migrated from Claude to Codex; GPT-5.4 cemented it as his daily driver. Metrics climbed: 200k WAU (Mar 19), built-in browser + doc/slide/sheet previews (April), 5M WAU (June, ~20% knowledge workers), and on July 9 Codex desktop became the new ChatGPT desktop app . Combined with ChatGPT for Work, WAU exceeded 40M. Designer Ed Bayes recalled leaving the all-hands realizing "the folks in the corner" now owned ChatGPT's future.

Growth Curves and Innovator's Dilemma

In a June 2026 Manhattan meeting, Embiricos (now heading enterprise product) drew Codex's growth curves on a whiteboard: cloud's brief spike flattening; post-GPT-5 CLI steeper; editor extensions higher; then 5.3 Codex + desktop app dwarfing all prior lines. Shipper saw Clayton Christensen's Innovator's Dilemma diagram — each compute shift (mainframe → PC → phone) spawns a simpler form factor and larger market. OpenAI compressed 50 years of that pattern into 10 months. Two factors let Codex escape the incumbent's curse: (1) deliberate parallel bets gave Codex independent iteration space outside ChatGPT's mass; (2) using Codex to build Codex changed product-development physics — a few people could ship new forms in weeks. Tibo: "Now it takes far less persuasion."

Future: Async and Cloud Return

The whiteboard held one more curve. Millions trust local agents; the next step revisits the original failed vision: an agent that keeps working after you close your laptop. Embiricos said they're pivoting to async — meaning a return to cloud. Shipper noted Embiricos's laptop sat on the table, lid shut tight.