ChatGPT

There is one thing to settle: whether you need a single assistant for everything.

Who it is for. For people who want one tool for email, study, images, document analysis, and voice: ChatGPT keeps text, voice, images, file analysis, and an agent in one window, while text chats are unlimited and free. For those already paying for Plus and wanting AI to do rather than advise: Work mode creates a spreadsheet, presentation, or website from your emails and files.

Who else. For those who need GPT-6 Astra for complex tasks involving files and software. Access is rolling out gradually; terms and a comparison with Sol are collected in the model card.

Who it is not for. Anyone who needs a strong free AI: one stands right next door, at DeepSeek.

Choose between ChatGPT and DeepSeek: the decision point is in section 7.

What vibecoding isStarting from scratch: how to give AI tasks with no coding experience

Since July 2026, ChatGPT has three levels in one window: Chat for questions, Work for finished files, and Codex for code.

Models. The chat runs GPT-5.6, Luna on Free and Go and Sol on the paid plans. On Sep 22, 2026 Work and Codex moved to the new GPT-6 generation, Sol and Luna, and the top GPT-6 Astra is there from Plus within a limit. Astra reaches the chat itself as GPT-6 Pro only from the Pro plan.

Chat. Questions, texts, web search, analysis of uploaded files and photos. On Plus there is an effort slider under the answer with Instant, Medium and High steps: the higher it is, the longer the model thinks, and since Sep 14, 2026 Instant no longer switches to reasoning on its own. Extra High and Pro open up from the Pro plan; on Free and Go there is a Think button on GPT-5.6 Luna instead of the slider.

Work. Give it a complete task, and the agent plans the steps, connects to Gmail, Slack, Google Sheets, and Notion, shows progress in the right-hand column, and delivers a document, spreadsheet, presentation, report, or website. It asks you to confirm important actions. Tasks can be scheduled or triggered by an event: an email arrives, a channel message appears, or a pull request changes.

To translate or summarize a page you already have open, consider the Sider browser extension. Its card covers model choice, credits, and sending page content to the cloud.

If your task is just a presentation, check out the Gamma editor: it builds slides from a topic or your own text and lets you edit individual elements.

The former ChatGPT Agent mode is no longer available. Its history and how it differs from the current ways of working are covered in a separate card.

Dots. Since Sep 29, 2026, you can hand a persistent assistant an ongoing job that continues between conversations. A dot runs on GPT-6 Astra, uses its own cloud computer and browser, keeps context and comes back on its own with a result or a question. You can message and call it in ChatGPT, and also talk to it in Slack and Teams. Cloud work keeps going even when your computer is off.

For example, you ask it to watch for changes in launch materials: it drafts edits to copy, slides and documents, and brings important decisions to you for review. A connected channel does not start monitoring on its own: say what to track, what result you need and when to notify you. For code, a dot can create a task in Codex.

First experience. Early access participant Matt Page described editing a video in Descript: the dot picked footage, added subtitles and carried edits across versions. The voice and creative decisions stayed with the author, and the result went through several reviews. This is his experience, published after the Sep 29 launch, not our own quality measurement.

Dots were announced Sep 29, 2026; features and limits checked against OpenAI on Sep 30, 2026. We did not run Dots on a task for this card.

Sites and Codex. Work publishes websites by link on the chatgpt.site domain, and Codex writes and edits code in the same app. Sites are available from the Plus plan, while Codex has limited free access.

Voice, images, memory. The voice mode on GPT-Live-1 listens and talks at the same time, searches the web and shows cards right in the conversation. Free and Go run a lightweight version, Plus gives up to three hours a day. ChatGPT Images 2.5 draws on every plan; the exact Flare and Sunburst variants can be picked through the API. In GPT Image 2.5 card we compare them on examples of generation and edits. Memory has updated itself since Jun 4, 2026, and the Sources button under the answer shows which memories shaped it.

What was actually done. In the official video, the agent created a supplier spreadsheet from a lamp sketch in two and a half minutes, then generated three product photos and published a store page. We did not run ChatGPT on a task for this card ourselves: the section's facts come from OpenAI's help center, changelog, and launch video.

You don't need to install anything.

Step 1. You can ask the first question without signing up. History, files and voice need an account: sign up by email or through Google, Apple or Microsoft. No phone number is asked at signup (checked Sep 5, 2026).

Step 2. Work and Sites become available after upgrading to Plus.

Apps. On iPhone and Android use the official apps: the developer must be listed as OpenAI.

Computer. On a computer the web version is the more reliable option. The desktop app for macOS and Windows installs from the download page.

Launch a dot. On an eligible plan, open Dots in ChatGPT in a desktop browser or create an assistant in the desktop app. Connect the services you need and give it its first responsibility. Once set up, the same dot is available in the mobile app with a supported update; mobile browsers are not supported. If the section has not appeared yet, check the plan terms below: access is rolling out gradually.

Two signals: the audience is record-breaking and market share is shrinking.

Audience. More than one billion people a week according to OpenAI's official post of Aug 31, 2026; in February, the company cited 900 million, with more than 50 million paying subscribers. Ads on free plans reached a pace equivalent to a billion dollars a year in under 200 days. The most crowded AI product in the world.

Market share. By web visits to chat assistants ChatGPT holds 55.5% for August 2026, Gemini 25.6%, Claude 9.3% (Similarweb). A year ago ChatGPT had 73.3%: the leader is still first, but it shares the market. Agencies use different methods, so keep the number together with its date.

ChatGPT's place among the assistants in our catalog is shown under the card title, and the full tool ranking is on the shared page.

You do not have to pay: text is free; you pay for a stronger model, Work, and limits.

Free entry. Free gives unlimited text chats on GPT-5.6 Luna, search, images and file uploads with limits, plus 500 MB for files. In Free and Go an ad can appear under the answer: by the end of September 2026 OpenAI shows them in more than 60 countries, from the US and Europe to Japan and India.

Paid tiers. Go at $8 raises the limits on files, images and voice. Plus at $20 gives GPT-5.6 Sol in the chat and the new GPT-6 Sol, Luna and Astra in Work and Codex within a limit, plus full Work, Sites and event-triggered tasks from email or Slack.

Pro and teams. Pro at $100 gives 5 times the Plus limits, the Sol Pro model and GPT-6 Astra in chat with a limit of 50 messages a week. The $200 tier with a 20x limit has been temporarily closed to new subscribers since Sep 10, 2026. Personal plans have no annual billing, and for teams of two or more there is Business with seats at $20 and $100 a month with annual billing.

For developers. Since Sep 22, 2026 the GPT-6 Sol API costs $2 per million input tokens and $10 per million output, half the price of GPT-5.6 Sol. The smaller GPT-6 Luna takes $0.10 and $0.50, the top GPT-6 Astra $10 and $50.

Who gets Dots. Under the terms as of Sep 30, 2026: Pro $100, $200 and $500 for users over 18 outside the European Economic Area, the UK and Switzerland. Business Premium and Enterprise get them worldwide; on Enterprise the assistant is off by default and an admin turns it on. This is a phased rollout, so having an eligible plan does not guarantee the section appears the same day.

The first dot is included in Pro and Business Premium at no extra cost. Chatting with it does not count against your ChatGPT limit, but there is a separate quota for deep work, expanded for the first month after launch. Tasks the dot launches or runs in Work and Codex use those products' regular limits. Free Codex and a Plus subscription do not unlock Dots on their own.

Four questions: does it train on your chats, where is the data stored, what about lawsuits, and how often does it go down?

If you choose based on how your chats are stored, check out Lumo privacy: its card covers chat history encryption, the model's access to your prompt, and opt-in feedback sharing.

Trains by default. On personal plans, OpenAI may use your chats to train models until you turn off "Improve the model for everyone" in settings; you can turn it off at any time, but this does not erase data already collected. Temporary Chat is not used for training and is deleted after 30 days. Business and Enterprise do not train on data by default.

Data in the US. Regional data storage is available only with Enterprise; personal plans run on OpenAI's infrastructure. The rule is simple: don't upload contracts, clients' personal data, or trade secrets to the chat.

Lawsuits and leaks. In the New York Times v. OpenAI case, on Jan 5, 2026 the judge confirmed the transfer to plaintiffs of an anonymized sample of 20 million conversations. In February 2026, Check Point found a flaw through which one malicious request could extract messages and files from the code-execution environment; OpenAI fixed it by Feb 20, 2026, and found no evidence of exploitation.

Noticeable outages. The official status feed lists 15 incidents in the 24 days before Sep 5, 2026, including a complete chatgpt.com outage on Aug 20, 2026 and a disruption on Sep 3, 2026 that drew about 37 thousand reports at its peak on Downdetector. Keep a backup tool for deadline-sensitive work.

Ads see context. On Free and Go, ads are selected by the conversation topic and, if personalization is enabled, by previous chats and memory; advertisers do not see the chats themselves. Separate September news: GPT-6 Astra safety checks may pause a task and ask for confirmation, as the vendor itself warns.

Stop a dot. In the assistant's profile, the Activity section shows child tasks and results. Pause stops the current main task, but not the whole set of assignments: stop child tasks separately in Activity, and cancel future runs in Scheduled. Ending a call is not enough either. Connecting a computer and services grants separate permissions; asking it to draft an email does not authorize it to send it.

The choice is not "who is smarter," but "how much access hassle and payment are you willing to deal with."

Different barrier. DeepSeek has a chat that is free in full, and the first question can be asked without signing up. ChatGPT gives voice, images, memory and Work mode, which the DeepSeek chat does not have.

Different API price. DeepSeek V4.1 Flash takes $0.15 per million input tokens and $0.60 per million output off-peak, GPT-6 Sol $2 and $10: a difference of 13 and 17 times. The top GPT-6 Astra costs $10 and $50, and for that money OpenAI gives its strongest model and an ecosystem of plugins and apps.

Who else is nearby. Gemini wins on a million-token context and on API price: Gemini 3.8 Flash costs $0.75 and $3.75 per million until the end of 2026. And Kimi brings back finished files for free.

In four years, the chatbot has become a workspace with an agent inside, and its audience has grown to a billion people per week.