Codex Major Upgrade: GPT-5.6 Powers Full-Stack Development from Idea to Deployment

Article

OpenAI's Codex has received a significant update. According to OpenAI's official demo video, more than 6 million developers use Codex every week. This update revolves around a new model, browser control, one-click deployment, task orchestration, and mobile experience — aiming to evolve Codex from a "code-writing assistant" into an "intelligent collaborator across the entire development workflow."

Background: OpenAI Codex was originally launched in 2021 based on the GPT-3 architecture, fine-tuned specifically for code generation tasks and serving as the foundational technology behind GitHub Copilot. Early Codex excelled primarily at converting natural language descriptions into code snippets, supporting dozens of programming languages including Python and JavaScript. After 2023, with the rise of the GPT-4 series, Codex underwent architectural consolidation, gradually evolving from a standalone API service into a component of a larger ecosystem. Its integration into ChatGPT with GPT-5.6 marks the completion of Codex's strategic transformation from "single-point code generation" to "full-pipeline development agent."

Codex Joins ChatGPT: A Unified Entry Point That Connects the Toolchain

The most notable change is that Codex has been officially integrated into the ChatGPT application, with its own dedicated section displayed alongside the new Work Agent. For developers, it's still the familiar Codex — but now it's deeply connected with ChatGPT conversations, Deep Research, and other capabilities, significantly reducing the cost of switching between tools.

OpenAI revealed that Codex has shipped more than 150 updates in the past two months. This high-frequency iteration pace signals that OpenAI is treating Codex as a core product that requires continuous refinement, not a one-time feature launch.

New Model GPT-5.6 and Ultra Mode

The foundation of this update is the new frontier model GPT-5.6 (codenamed Sol in the demo), available to all users and designed for extended deep reasoning on complex problems. Combined with the newly introduced Ultra Mode, the model gains a reasoning budget far beyond normal limits — particularly suited for tackling those long-shelved, high-difficulty projects.

Notably, OpenAI is candid about one tradeoff: Ultra Mode burns through your token quota faster. Greater reasoning power means higher usage costs, and developers will need to weigh that against task complexity.

In Ultra Mode, Codex automatically splits tasks across multiple sub-agents running in parallel, with no manual intervention required. Developers can follow up on individual sub-tasks or open the panel to see the full "team" working in concert. One goal issued, multiple workstreams running simultaneously — this is a pivotal step for Codex toward "agent orchestration."

Multi-Agent Architecture Explained: Multi-agent architecture is one of the most cutting-edge paradigms in AI engineering today. The core idea is to decompose complex tasks into multiple sub-tasks, handled in parallel by independent agent instances, then consolidated by an orchestrator. This approach draws from the MapReduce concept in distributed computing: a master agent handles task decomposition and scheduling, sub-agents handle execution, and the master agent aggregates the final output. The parallel execution mechanism in Codex's Ultra Mode essentially hides a complete task scheduling system behind a single user instruction. The tradeoff is that each sub-agent independently consumes reasoning tokens — which is precisely why Ultra Mode burns through tokens faster.

Browser and Computer Control: Letting AI Actually "Get Its Hands Dirty"

In the demo, engineers specifically emphasized that browser and computer control may be the most anticipated feature of this release — and GPT-5.6 makes the entire process dramatically faster.

Computer Use Technical Background: Browser and computer control (Computer Use) has been one of the core frontiers that major LLM providers have been racing to break through since 2024. Anthropic was first to the punch in October 2024 with Claude's Computer Use feature, allowing AI to directly control the mouse, keyboard, and screen. The technical approach typically combines a Vision module with an Action module: the model interprets the current interface state from screenshots, then outputs structured operation commands (such as click coordinates and keyboard input). OpenAI's implementation builds on this with deeper integration into the development toolchain — particularly the "app shot" concept seen in simulator screenshots, which simultaneously captures the visual frame and runtime context (such as the current route and component tree state), providing far richer semantic information for subsequent code modifications than a plain screenshot would.

A typical scenario is iPhone app development. On a Mac, developers can use a keyboard shortcut to capture an "app shot" from the simulator — a smarter capture method that records both the screen and the app's runtime context simultaneously. Then, with a single instruction like "walk through each screen and capture App Store screenshots in English and French," Codex can control the simulator and perform the clicks automatically, running entirely in the background while developers work on other things.

Click Annotation and Inline Code Editing

For web applications, Codex's built-in in-app browser is equally powerful. Developers can run a freshly generated app directly in the browser, and if they need to tweak something, they simply click and annotate the specific element. Codex captures the visual context and makes precise changes accordingly.

The browser also adds support for apps that require login, including passkey authentication. Developers can also switch to a DOM view where they can not only inspect elements but also edit code lines inline. The ability to switch freely between visual manipulation and direct code editing brings the experience closer to "pair programming."

Sites: One-Click Deployment from Zero to Full-Stack Application

After writing code, "how do I get this online" is often the next hurdle. That's precisely where Sites comes in. Previously, Sites was only available in the enterprise tier for internal team sharing; this update opens it to all users, supporting the direct publication of any web app built with Codex.

Developers simply say "publish to Sites" and get built-in hosting, authentication, a persistent database, and even file storage. From an idea to a genuinely usable full-stack application — with virtually no infrastructure or deployment details to worry about.

BaaS/PaaS Technical Background: The Sites feature represents a further convergence of "Backend as a Service (BaaS)" and "Platform as a Service (PaaS)" philosophies. Traditionally, getting from code to production required developers to handle server configuration, DNS resolution, SSL certificates, database initialization, authentication integration, and a host of other DevOps tasks — which is exactly why platforms like Heroku, Vercel, and Netlify rose to prominence. Sites abstracts all of that away, with built-in authentication (similar to a managed Auth0 service), a persistent database (similar to a managed PostgreSQL), and file storage (similar to S3) — all without writing a single line of infrastructure configuration. This "zero-ops" model combined with AI code generation has the potential to compress the end-to-end time from natural language description to a live, accessible web service down to minutes.

This "zero-ops" deployment experience dramatically lowers the barrier to launch for independent developers and small teams, and extends Codex from a code generator into a complete application delivery platform.

Task Orchestration: Managing Multiple Projects While Keeping Context Clear

As developers delegate more and more work to Codex, maintaining clear context across multiple parallel projects becomes a new challenge. To address this, Codex can now proactively manage its own tasks on your behalf.

For example, a single instruction like "find five bugs in Linear, create a separate task and isolated work tree for each, and flag the most critical ones" can simultaneously launch and track all the work from a single coordinating thread, with cross-references between tasks.

Git Work Tree in Practice: The "isolated work tree" mentioned in task orchestration refers to git worktree, a relatively advanced Git feature that allows developers to check out multiple different branches from the same local repository into separate directories simultaneously, without interfering with each other. This is especially valuable when handling multiple parallel bug fixes: the traditional approach requires frequent branch switching (git checkout), which can cause conflicts with uncommitted changes. The worktree approach gives each task its own working directory and filesystem state, allowing sub-agents to truly work in isolation in parallel. By automating this advanced Git feature, Codex lets developers enjoy the benefits of parallel multi-tasking without having to manually manage the creation, switching, and cleanup of work trees.

What's more, work doesn't have to originate in Codex. If inspiration comes from a ChatGPT conversation or conclusions drawn from Deep Research, you can use @mentions to bring the full context into Codex — no more manual copy-pasting.

Development That Keeps Moving on Mobile

The ChatGPT mobile app received several meaningful upgrades in this update. Three new capabilities are particularly highlighted:

• Create, search, open, and manage tasks directly within conversations
• Stronger change review capabilities, with flexible filtering between staged, unstaged, and branch states, or direct branch diff comparisons
• Direct SSH host connections (e.g., remote dev machines) to initiate new tasks from your phone

This means developers can kick off a task on their desktop, track progress on their phone at any time, and jump back in whenever needed — a truly continuous development experience that doesn't fall apart when you step away from your desk.

Seamless Pull Request Workflow

Codex has also integrated tightly with the PR workflow that sits at the heart of daily development. It can detect the working branch for each task, and when a branch has an open PR, the summary panel automatically surfaces the relevant information. Developers can follow up, address issues, and complete merges — all without leaving the app.

The Deeper Value of PR Workflow and CI/CD Integration: The Pull Request workflow is the core mechanism of code collaboration in modern software engineering, originating from the feature GitHub introduced in 2008 and now an industry standard. CI (Continuous Integration) checks typically include automated testing, code style checks (lint), security scanning, and build verification — a failed check blocks the merge. The deeper value of Codex's PR integration lies in the fact that it has full task context: which files were changed, why they were changed, and what the original requirements were. This gives it a significant advantage over the traditional "read the error message from scratch" approach when fixing CI failures — the model can reason by combining build logs with the intent behind the code changes, rather than blindly guessing based on error messages alone. This represents a critical capability leap for AI assistants evolving from "code generation" to "engineering intelligence."

If a CI check fails, you can ask Codex to locate and fix the issue directly — with full context already in hand. The new Pull Requests tab lets you browse all code changes, leave inline comments, complete full code reviews, even mark your own PR as ready and complete the merge — all without navigating to github.com, keeping developers in a productive flow state.

Summary: Codex Is Becoming a Full-Cycle Development Agent

Looking at this update as a whole, Codex's product positioning is now quite clear: it's no longer just a code completion or generation tool. It's a development agent capable of deep reasoning, browser control, parallel task orchestration, one-click deployment, and autonomous PR management. GPT-5.6's extended reasoning and multi-agent parallel execution capabilities are the technical engine powering all of this.

For developers, the most noteworthy aspect is the "full-pipeline integration" — from idea (ChatGPT/Deep Research) to coding (inline editing), running (built-in browser), deployment (Sites), and collaboration and review (task orchestration, PR management). Codex is attempting to consolidate the entire development pipeline into a single unified entry point. Whether this integration will truly boost productivity still needs to be validated in real-world projects — especially since the token costs of Ultra Mode require careful consideration. But one thing is certain: AI programming tools are rapidly evolving from "helping write code" to "acting as the agent for the entire development cycle," and this Codex update is the most direct manifestation of that trend.