ChatGPT Integrates Codex: A Full Hands-On Breakdown of Work and Codex on Desktop

ChatGPT Absorbs Codex, Desktop App Becomes an AI Command Center

ChatGPT recently completed a major product integration: folding the standalone code tool Codex into a redesigned desktop app. Open the new application and you're no longer greeted by the familiar chat window — instead, you're met with two core modules: Work and Codex. Chat, coding, image generation, writing, and local computer tools are all unified under a single entry point.

The Technical Evolution of Codex: Codex was originally released by OpenAI in 2021 as a GPT-derived model specifically fine-tuned for code generation, and served as the underlying engine for GitHub Copilot. By training on billions of lines of public code, it learned to interpret natural language instructions and generate code snippets across multiple programming languages. Its integration into the ChatGPT desktop app represents a strategic shift from "code completion tool" to "engineering execution agent" — Codex is no longer just a coding assistant, but an execution engine capable of orchestrating tools, managing files, and driving complete engineering workflows.

This integration is far more than a simple feature stack. In hands-on testing, the desktop version proves significantly more capable than the web version when it comes to local functionality: it supports a skill system that can automatically execute tasks across apps, access local files and applications, includes a built-in terminal, and offers stronger data privacy guarantees. The results were immediate — desktop daily active users grew by one million in a single day, with local capabilities as the primary driver.

Interestingly, the original chat assistant hasn't disappeared — it's been tucked into the sidebar, opening as a small window in the bottom-right corner for quick, ad-hoc conversations. A thoughtful design decision: regular chats only consume Chat quota and don't draw from the Work and Codex agent quota.

The Design Logic Behind Agent Quota Separation: Work and Codex use an Agent Quota that is separate from regular Chat, reflecting the engineering realities of AI resource allocation. Agent tasks require multiple rounds of tool calls, sustained context maintenance, and longer reasoning chains — consuming far more compute than a single conversation. A 13-minute PPT generation task, for instance, may occupy context windows and trigger tool calls equivalent to hundreds of ordinary chats. Separating the quotas both protects resource availability for intensive engineering tasks and prevents users from exhausting their agent allowance through frequent casual chats — a refined design that balances user experience with compute costs.

Work: From "Answering Questions" to "Delivering Results"

ChatGPT Work is not a simple upgrade to the chat function — it's an entirely new work execution system. It automatically gathers context, completes tasks across workflows, and genuinely bridges the gap from "answering questions" to "delivering results directly."

A typical office scenario demonstrates its capabilities: asking Work to generate a reusable PPT template. The process unfolds in two steps — first, calling Presentations to create and review the slide deck, then using Template Creator to package it as a reusable template. The task took approximately 13 minutes, during which the agent automatically invoked an open-source skill called「小黑配图」, applying consistent illustration styles, color schemes, and annotation elements across a 13-page template.

The most important detail: the template isn't just saved to the local template directory — it's also created as a standalone, default-enabled skill. Future use is flexible: select it directly from the template library, or trigger it instantly with a slash command.

The Paradigm Shift from Prompt Engineering to Skill Systems: The "skills" described here are essentially packaged tool-use chains, similar to Agent tools in LangChain or plugins in AutoGPT. After a user demonstrates a complete workflow once, the system codifies the steps, tool choices, and parameter configurations into a reusable skill module — no need to re-describe the task next time. This marks a shift in AI interaction paradigms: from "carefully crafting prompts each time (Prompt Engineering)" to "demonstrate once, reuse many times" programmatic workflows. Regular users no longer need to master complex prompting techniques — they just demonstrate the process once, like training an apprentice, and the AI converts it into a standardized procedure.

Template Reuse: Professional Output, On Demand

With the template ready, Work was asked to generate a "Q2 2026 East China Coffee Store Operations Analysis Report" (10 to 15 pages). The task ran for about 11 minutes, delivering charts, data cards, and layouts unified under a consistent visual style — all in editable PPTX format for easy follow-up adjustments.

Although the demo used simulated data and content depth was limited, the level of completeness was already impressive. This is the core value of Work: consistent reuse when uniformity matters, rapid generation when customization is needed — professional output, ready the moment you ask.

Codex: A Fully Automated Pipeline for AI Video Production

If Work targets everyday office tasks, Codex is the dedicated partner for complex engineering work. One of its more advanced use cases: using custom skills and code to automatically produce a complete video from scratch.

This "one-stop video production skill" covers the full pipeline: topic selection, scripting, storyboarding, audio, and video rendering. The topic selection stage generates 10 candidate topics and recommends three directions; in the scripting stage, the system doesn't just write a draft by feel — it first performs searches and fact verification around the selected topic.

Only after fact-checking is complete does the system generate the narration script and break it down into individual shots. A 5-minute Chinese-language video ultimately produced around 1,100 characters of narration, split across 22 shots. The storyboarding stage also includes a "confirmation gate" requiring human sign-off on shot accuracy, aspect ratios, and illustration style — ensuring the project stays on track.

From Illustration to Voiceover to Rendering: Engineering Thinking in the Details

In the illustration phase, the system generates complete image prompts for each of the 22 shots, locking in character design, line style, color meanings, and white space rules to ensure a unified visual language throughout the video. After roughly 30 minutes, 22 body images and 3 thumbnail images were delivered. One caveat: unwanted English labels may appear in the upper-left corner of generated images — this can be fixed by updating the skill with an additional constraint.

The voiceover stage revealed the importance of permission management: restricted network environments can block voice generation. It's recommended to set permissions to "Full Access" at the start of the task to avoid approval interruptions mid-workflow.

The final video rendering shows genuine engineering ingenuity: Codex invokes Python code to dynamically set each shot's duration based on the actual length of its corresponding audio segment, then assembles everything into a complete video.

The Technical Logic Behind Dynamic Duration Video Rendering: Codex calls a Python script to achieve audio-video synchronization, essentially using open-source libraries like MoviePy and FFmpeg to read the actual duration of each TTS (text-to-speech) audio clip, then dynamically set the display duration of the corresponding image frame, before stitching everything into a complete video stream. This "audio-driven timing" approach is more natural and fluid than a fixed frame rate, eliminating the awkward gaps where narration has ended but the frame is still on screen — or where the image has already cut while the voiceover hasn't finished. More importantly, this logic is packaged as reusable code rather than being re-inferred by the language model each time — and this is precisely the key engineering advantage that sets Codex apart from ordinary AI conversation: determinism, high efficiency, and zero Token waste.

Critically, the final deliverable isn't just an MP4 — topic selection, script, storyboard, image prompts, fact notes, per-shot audio files, various manifest files, and a final acceptance report are all preserved in a clean, organized directory structure.

Work vs. Codex: One Gets Things Done, One Writes Code

The two modules may seem similar — so why split them? One-sentence summary: go to Work for daily tasks and office work; go to Codex for code and engineering.

Work is the "task partner" — great at completing office work across files and applications, organizing and generating documents, spreadsheets, images, and presentations, with a focus on delivering results directly. Codex is the "engineering partner" — focused on handling complex tasks through code and skills. The key difference lies in how they execute: for a final step like video synthesis, Codex invokes reusable code rather than asking the language model to reason through it again each time — the latter is both slower and more Token-expensive.

Why "Reusable Code" Matters More Than "Reasoning Every Time": The inference process of a large language model (LLM) is fundamentally probabilistic sampling — the same instruction can produce slightly different outputs each time. For engineering tasks like video rendering that demand precise, deterministic results, encoding the logic as code means zero ambiguity, full auditability, and easy debugging. This is also the deeper reason Codex is positioned as an "engineering partner" rather than a "chat assistant": in production scenarios that require reliable delivery, the determinism of code is far more valuable than the flexibility of a language model.

The significance of this integration goes well beyond "merging two apps." It signals that AI tools are shifting from conversational interaction to result delivery — moving away from dependence on carefully crafted prompts toward skills that can be reused after a single demonstration. For creators and knowledge workers, the barrier to AI automation is dropping dramatically. A locally-powered intelligent workstation is already taking shape.

Key Takeaways