ChatGPT for marketers: features, plans, and working scenarios

OpenAI renames things more often than anyone: in 2026 GPT-4o, GPT-4.1, and o4-mini left the app, then GPT-5 itself, and in August the familiar choice between Instant, Thinking, and Pro modes disappeared too. The current logic is one main model and a slider for how long it thinks.

• GPT-5.6 Sol – the flagship for paid plans. Above the input field sits a Thinking slider with four positions: Instant for quick answers, then Medium, High, and Extra High for tasks that need a chain of reasoning. You no longer pick a model, you pick the depth.
• GPT-5.6 Luna – the fast, lighter model for Free and Go. Since August 2026 plain text chats on it have no cap; a Think button handles harder questions, but it also runs on Luna, not Sol.
• GPT-6 Astra – the new generation, announced on September 4, 2026 and rolling out to Plus, Pro, Business, and Enterprise. The API version has a one-million-token context; in the app the limits are smaller and depend on the plan. Pro subscribers get Astra Pro.
• GPT-6 Sol and GPT-6 Luna – released on September 22, 2026 as cheaper successors to the 5.6 versions: first in the API, Codex, and the ChatGPT Work mode on paid plans, with GPT-6 Luna available to Free and Go users in the desktop app. In the main chat the switch from 5.6 is gradual, so both generations show up in the app for a while.
• Separate models for media and agents: GPT Image for pictures, a dedicated voice model, Codex for code. All are called from the same window.

A practical rule: Instant where a mistake is cheap, Medium or High for documents and calculations, Extra High once a day for the task you would hand to an analyst. The context window in the app is narrower than Gemini's: on Plus roughly 54,000 tokens in fast mode and 256,000 in reasoning mode, more on Pro. A hundred-page brief is better uploaded as a file into a project than pasted as text.

Prices below are for the US at the time of publication (September 2026) and differ by country. OpenAI now describes limits as limited, expanded, and maximum access and no longer publishes exact counts for Deep Research or images, so the table is qualitative on those too.

For a marketer the working choice is Plus. Go looks cheap, but without Deep Research and the agent it is a chat with ads, and it does not hold up as a work tool. Pro pays off if you run Deep Research and agent tasks daily or hit the Plus limit before lunch. For a team of two or more, Business on annual billing costs the same as Plus per person and adds shared projects and data protection.

Deep Research is a mode where the model spends 5–30 minutes browsing dozens of sites, cross-checking sources, and assembling a report with numbered citations. Before it starts it asks clarifying questions: region, period, what counts as a competitor. In marketing that means a market overview before entering a new city, a competitor breakdown in a location, or facts for an article with links instead of "according to experts".

Agent is the mode where ChatGPT gets its own browser and does what you would do by hand: opens sites, clicks, fills forms, downloads exports, builds a table. In July 2026 OpenAI announced the end of its standalone Atlas browser, and agentic browsing moved into the desktop app and a Chrome extension. Available on Plus and above. I use it for chores like "go through 15 developer websites and collect price per square meter and installment terms": 20 minutes of model time against half a day of an intern's.

ChatGPT Work is an agent for longer tasks that end in a finished file: a report, a spreadsheet, a deck. It pulls data from connected apps and files, not only from the chat. A quarter of CRM exports plus a media plan turns into a summary deck that only needs a check and a signature.

Projects are folders with a standing instruction, files, and their own chats. A project has separate memory: since August 2026 it can be switched to project-only mode, so chats about one client do not pull context from another. It is the direct counterpart of Projects in Claude and Gems in Gemini. I keep a project per property: audience, prices, floor plans, banned phrases, copy that worked.

GPTs are saved versions of ChatGPT with an instruction that can be published for others. For a long time this was the main way to hand a team "our copywriter" or "our analyst". In September 2026 OpenAI announced a gradual retirement of GPTs in favor of plugins: the same instructions plus connected apps. Dates are set for enterprise workspaces, not yet for personal plans. I build new things in Projects.

Memory lets ChatGPT keep facts about you across chats: what you do, which tone you prefer, which projects you run. In 2026 the system was rebuilt to periodically re-synthesize what it knows instead of hoarding old notes. For a marketer this is convenient and risky at once: a detail from one client's chat can surface in copy for another. Keep client work in Projects with isolated memory.

Canvas is a workspace next to the chat where text or code is edited in parts: select a paragraph, ask to shorten it, the rest stays untouched. For a landing page or a long email it beats a dialog where every edit returns the whole text.

Images come from the GPT Image model right in the chat: from a description, from a reference, or as an edit of an uploaded picture. Its strengths are text inside the image and targeted edits without rebuilding the whole scene. Upload a facade render, ask for "evening, warm light in the windows, people at the entrance", and get a variant with the geometry preserved. Free gives a few images a day, Plus gives dozens per few hours.

Video is gone from ChatGPT. OpenAI announced the closure of Sora on March 24, 2026, switched off the app on April 26 and the API on September 24, and has no plans to build video generation into ChatGPT. For clips that leaves Veo inside Gemini and specialized services; for avatars and dubbing I still use HeyGen.

Voice mode is a spoken conversation with interruptions and camera input. For a marketer it is a way to dictate a brief on the way to a meeting or rehearse a pitch: the model asks the questions a client would.

Connectors and apps link Google Drive, Gmail, SharePoint, Dropbox, Notion, GitHub, Slack, and hundreds of third-party apps such as Canva or Figma. On Business and Enterprise the Google and Microsoft connectors can write as well as read: create an event, send a draft, update a document. In substance this is what MCP does in Claude, delivered through OpenAI's catalog. Codex is a development tool included in Plus and above; a marketer needs it only for small scripts such as merging two ad-account exports.

1. 01Competitor price monitoring. Once a week the agent goes through the project websites in a location and builds a table: price per square meter, offers, installment terms, new phases. Deep Research adds context from news and reviews. Half an hour of checking instead of a day.
2. 02Creatives for tests. From one render GPT Image builds a series per audience: family with kids, investor, couple without children, relocating buyer. Six variants in 20 minutes, an ad test in a day, production joins for the winner.
3. 03Reading reports. A quarter of search-ad and CRM exports goes into a project with one question: in which channels did cost per meeting grow faster than cost per inquiry, and why. The model reads tables in full and draws the charts itself.
4. 04Scripts and objections. A month of sales call transcripts turns into a list of objections with frequency and quotes, then into landing page arguments and an updated script. ChatGPT is good here; the final edit in the brand's tone goes to Claude.
5. 05A deck for the weekly meeting. ChatGPT Work takes the media plan, the export, and last week's deck from Drive and assembles a new one with updated numbers. What remains is checking the figures and cutting spare slides.

The scenarios work when each has an owner, a schedule, and a human check. I wrote about building AI into the process rather than into a pile of chats in the article on AI tools for marketers, and about multi-step automation in the piece on AI agents in marketing. At Matveo we assemble such setups for developers and agencies as part of the marketing loop: the agent collects the data, a person makes the decisions.

ChatGPT is officially available in about 170 countries, and the list is worth checking before you buy, because features and ads differ by market. Russia is the notable gap: OpenAI has blocked Russian IP addresses since mid-2024, and since September 6, 2026 the chatgpt.com domain has also been cut off by Russian internet providers. Cards issued by Russian banks have been rejected by the payment processor since 2022.

Teams that serve the Russian market from abroad use the same setup as for Claude and Gemini: an account tied to a foreign phone number, a VPN with a dedicated IP, and a foreign or virtual card. Two caveats. OpenAI recognizes the address ranges of popular VPN providers, so only a private server or a dedicated-IP service works. And such workarounds breach the terms of use, so the account can be closed along with the paid month; keep working documents outside ChatGPT as well.

The takeaway for a working stack: ChatGPT is the generalist with the strongest agent and the largest app ecosystem, Claude covers text and documents, Gemini covers research, visuals, and Google Workspace. A solo marketer gets by with ChatGPT alone or with Claude plus Gemini; three subscriptions only make sense for a team where each tool has an owner.

• Facts without sources. Outside Deep Research the model answers from memory and confidently quotes wrong market figures and residential projects that do not exist. Any number only with a link.
• Marketing copy. Grammatically clean, but the same constructions repeat: triple lists, contrasts, sweeping claims. A draft, yes; the final version after an editor or through Claude.
• Names and plans keep changing. In one year OpenAI renamed Team to Business, removed the model picker, closed Sora and Atlas, and added then paused Pro tiers. Check everything here about money and models on the pricing page on the day you buy.
• Memory across projects. Without project-only memory, details from one client can leak into copy for another. Check the settings before uploading client data.