ChatGPT

ChatGPT is the assistant that made this whole category mainstream, and it remains the default for most people who use AI at all.

It has also stopped being only a chat window. OpenAI's own overview page now leads with "chat, work & code all in one place", and Codex — the coding agent — sits in the top navigation beside Business and Pricing.

This review covers the features that actually change how you work, the structural tools most people never touch, what happened to Sora, and the failure modes worth knowing before relying on it.

ChatGPT: The Essentials

• What it is: the most widely used AI assistant, covering text, images, documents, voice, web browsing, and code in one product.
• The positioning shifted: OpenAI now sells it as "chat, work & code", with the Codex coding agent given its own place in the navigation.
• Structure matters more than most people use: Projects hold files and instructions across conversations, and Tasks let it run on a schedule.
• Agent behaviour is built in — it can take multi-step actions with approval checkpoints rather than only answering.
• Video is gone. OpenAI discontinued Sora in 2026; the app closed in April and the API sunsets in September.

ChatGPT Capability Overview

Beyond The Chat Box

The single biggest gap between casual and effective use is structure. Most people open a blank conversation every time and re-explain their situation; the tools that fix that are already there and largely ignored.

Projects Carry Your Context

A Project groups related conversations under shared instructions, uploaded files, and a memory that persists across all of them. Set up once, and every chat inside it starts already knowing your context.

For anything you return to — a client, a codebase, an ongoing piece of research — this is the highest-return ten minutes available in the product.

Tasks

Tasks let ChatGPT run on a schedule rather than only when prompted. A weekly summary, a recurring check, a morning briefing — work that happens without you opening the app.

It moves the assistant from something you consult to something that produces on its own.

Custom GPTs

Custom GPTs package instructions and files into a reusable assistant you can share. They suit a repeated task with a defined method more than they suit general conversation.

Agent Behaviour And Its Guardrails

ChatGPT can take actions rather than describe them — navigating sites, filling forms, comparing options, and working through multi-step jobs.

The important design detail is the approval checkpoint. It pauses before consequential actions rather than proceeding on its own judgement, which is the difference between a useful agent and an alarming one.

Treat that pause as the feature it is. An agent that never asks is faster right up until the moment it does something you would not have authorised.

The question with any agent is not what it can do unsupervised. It is what it will stop and ask about.

Codex And The Coding Side

Codex is OpenAI's coding agent and now has its own prominence in the product navigation. It works across editors, the command line, and mobile, taking on tasks rather than completing lines.

This puts ChatGPT into direct competition with dedicated AI coding tools rather than adjacent to them. For developers already paying for ChatGPT, it removes a reason to add another subscription.

For teams deep in a specific codebase, dedicated editors still tend to win on repository awareness. The comparison is worth running on your own code rather than on benchmarks.

What Happened To Sora

This comes up constantly, so it is worth stating plainly: OpenAI discontinued Sora, its video generation model, during 2026.

The announcement came in March 2026, the app and web experience shut down on 26 April 2026, and the API sunsets on 24 September 2026. There is no announced successor and no timeline for video generation returning.

Reporting attributes the decision to economics rather than capability — the model was expensive to run and usage fell away after the initial novelty — alongside unresolved copyright and likeness questions. If you need AI video, you need a different vendor.

Habits That Change The Output

A few habits separate people who get a lot from ChatGPT from those who conclude it is overrated.

1. Create a Project for anything you will return to more than twice, and put your standing context in it.
2. Give it the actual document rather than your summary of the document — pre-summarising discards the detail that makes answers good.
3. State what you want the output to be for, not just what you want it to say. Audience changes everything.
4. Ask it to identify what is weakest about its own answer before you act on it.
5. Use Tasks for anything recurring rather than remembering to ask each week.
6. Verify anything factual that matters. It will state a wrong date with the same confidence as a right one.

Where It Still Gets Things Wrong

Confident errors remain the core risk. ChatGPT does not signal uncertainty reliably, so a fabricated citation, a wrong date, or an invented function name arrives in the same tone as a correct one.

Consistency varies. The same question can produce a strong answer and a shallow one on different days, which makes it hard to build a dependable process on top of.

Long conversations drift. Detail from early in a thread gets deprioritised, and the fix is starting fresh with a summary rather than continuing to push.

And the product surface is now large enough that most users never find half of what they are paying for — Projects and Tasks in particular.