APREX · EST. 2026
Since ChatGPT launched in November 2022, the text conversation has been the very foundation of OpenAI's consumer product. Now that fundamental structure appears to be changing — and it is code, not chat, that is driving the shift.
From code completion to autonomous agent
The original Codex, launched in July 2021, was a 12-billion-parameter version of GPT-3 fine-tuned on 159 gigabytes of Python code sourced from GitHub. The tool functioned as an advanced autocomplete — useful, but passive.
The new Codex is something fundamentally different. According to Stratechery and technical background information from OpenAI, the platform is now powered by codex-1, a specialized version of the reasoning model o3. Where the old Codex matched patterns, the new version can plan and iterate through complex, multi-step programming tasks on its own.
Where the old Codex matched patterns, the new version plans and iterates through complex tasks on its own
The agent operates in isolated virtual computing environments, can execute code, connect to GitHub, and handle multiple parallel programming tasks simultaneously — without requiring the user to hold its hand along the way.
How the agent loop works technically
At the core of Codex is what is described as an "agent loop": user input leads to model inference, which in turn can result in a tool call — for example, running a shell command. The result is added to the context, the model is queried again, and the cycle repeats until a final response is ready.
To handle the broad range of user surfaces — CLI, VS Code extension, web app, and macOS app — OpenAI has developed a bidirectional protocol called "Codex App Server". This separates the agent's core logic from the various client interfaces through a single stable API layer.
ChatGPT as a super-app for developers
With the integration completed in July 2026, ChatGPT is no longer primarily a conversational partner — it has become a coordination point for parallel, agentic development work. Engineering teams can now run simultaneous coding tasks across projects directly inside the ChatGPT app, with built-in cloud environments.
This raises the question Stratechery poses in its analysis: is OpenAI in the process of abandoning the very chat format it helped popularize? The answer is not obvious. It is more accurate to say that chat is being demoted from being the product's purpose to becoming one of many entry points within a broader agentic platform.
Reasoning as the foundation
The technical pivot is not just about functionality, but about underlying model architecture. OpenAI's reasoning models — o1 arrived in September 2024, o3 in April 2025 — generate long internal chains of thought before producing a final response. That makes them well-suited for tasks requiring planning rather than rapid text generation.
The fact that codex-1 is a specialized variant of o3 is no coincidence: reasoning and iteration are precisely what is needed to write code that actually works in complex systems. Code completion can be handled with pattern recognition. Autonomous software development cannot.
What this means going forward
OpenAI has not officially announced that it is "abandoning chat" — and that would be an oversimplification. But the integration of Codex into ChatGPT clearly signals the direction in which the company sees the greatest value. The platform is now being optimized just as much for doing things as for talking about them.
For developers and engineering teams, this is a concrete and immediate benefit. For users who rely on ChatGPT for writing, research, or everyday assistance, the change is for now mostly architectural — but it sketches out a product that is in the process of becoming something quite different from what was launched nearly four years ago.
Sources: Stratechery (Ben Thompson), OpenAI technical documentation via research