OpenAI’s GPT‑5.6 Arrives, Codex Disappears: A Hands‑On Review of the New ChatGPT

ChatGPT Desktop Client – Layered Workbench

OpenAI merged Codex into the ChatGPT desktop client and released GPT‑5.6. The new model consumes the 5‑hour quota very quickly (the author’s limit was exhausted after two tasks) but provides deeper reasoning and about 1.5× faster response times.

Four Stacked Layers

Chat : Traditional Q&A, brainstorming, and quick conversations. Available on web, mobile, and desktop.

Work : An agent that can access apps, files, and workflows, then generate tables, slides, documents, or web applications. It can decompose a large project into hour‑long independent steps, effectively acting as an AI project manager.

Codex : Software‑development engine for writing code, debugging, running tests, reviewing PRs, and coordinating multi‑repo work. Desktop‑only.

Sites : Public‑beta feature (Pro users) that turns a description, content, and data into an interactive website or lightweight app, with one‑click publishing and automatic syncing when underlying information changes.

1. Chat – Local and Web Sessions Linked

The desktop client embeds the web version of ChatGPT in a local window. Users can click “Add to task” in a web session to inject that conversation’s context into a local Codex project. The reverse works, but the UI requires clicking the “More details” button rather than asking directly in the side chat.

New UI elements include SunAgent and an annotation shortcut (select text → add to conversation).

2. Work – Turning Goals into Finished Materials

Work is defined as an agent that can:

Grab context from apps, files, workflows .

Generate tables, slides, documents, web apps as finished deliverables.

Decompose complex projects into small steps that run independently for hours.

Official example: a single request can convert customer research into a campaign brief, generate marketing assets, and adapt content for multiple markets while preserving context throughout.

OpenAI reports >5 million weekly Codex users, >1 million of whom use Codex for non‑software tasks, indicating the tool’s expansion beyond pure programming.

3. Sites – One‑Click Website Creation

Sites is in public beta for Pro users. The workflow consists of six steps:

Describe the desired website.

Add content, files, data, links, and constraints.

Preview directly in ChatGPT.

Iteratively request modifications.

Publish to obtain a URL.

If underlying data changes, ChatGPT can automatically update the site.

Creating a site after a few dialogue turns often yields better results than starting from a blank description. Sites can be set to private or shared publicly.

4. Layered Architecture Summary

Chat (question‑answer layer)
   ↓
Work (office delivery layer, long research + finished materials)
   ↓
Codex (engineering execution layer, long tasks + multi‑repo + PR workflow)
   ↓
Sites (publishing layer)
   ↓
Plugins (workflow wrappers) = Skills + Apps + Templates

The stack mirrors the evolution of IDEs: from plain‑text editors to full‑featured development workbenches.

5. Enterprise Case Studies

Virgin Atlantic

Input a structured customer‑journey path.

Work researched competitor processes and generated a dataset highlighting strengths, weaknesses, and future investment areas.

Compressed weeks of competitive analysis into a few hours, dramatically shortening insight‑to‑decision cycles.

Zapier

Traditional review of each lead took 35‑45 minutes.

Work built a QA/QC and analysis system, tracked conversion journeys, visualized contact‑point graphs, identified churn nodes, and aggregated results into a weekly executive dashboard.

Result: identification and delivery of a seven‑figure sales pipeline each month.

NVIDIA (GTC Conference)

Manual preparation consumed ~40 % of the conference timeline.

Work automated the entire workflow, enabling global team sharing and regional adjustments.

Automation now runs twice weekly inside ChatGPT, freeing time for strategic planning and customer success.

Shopify

Extracted Slack and project context daily into a “second brain”.

Custom skills turned the extracted data into follow‑up tasks.

Deployed across 3,500 non‑engineering employees to run research projects, identify high/low adoption groups, coordinate interviews, analyze transcripts, and extract reusable patterns.

RingCentral

Previously consolidated data manually across Salesforce, Jira, emails, bug reports, and chat.

Work now surfaces blockers, owners, and action items before meetings, scaling follow‑up from 6 pilot customers to about 80.

All teams stay aligned on execution.

The new ChatGPT desktop client thus provides a layered productivity platform that spans from simple chat to full project execution and publishing.