ChatGPT and Codex, reviewed
OpenAI's assistant has 1.2 billion weekly users and a coding agent that now works in the cloud. What is good, what the tiers cost, and what to watch.
ChatGPT is OpenAI's assistant, used by 1.2 billion people a week. Next to it sits Codex. OpenAI's head of product says it was first built for “a forgiving technical audience”. Today it is an agent that writes software on a computer, from a phone or in the cloud.
September brought a new generation of models, a developer conference and a long list of features. Here is how ChatGPT and Codex serve a company as of October 2026: what they are, what is genuinely good, what is not, and our verdict.
What they are now
ChatGPT has become more than a chat. Work, launched as part of the Plus plan, carries out tasks of several steps with the files, apps and tools a person is allowed to use. Its purpose, in the words of that head of product, Thibault Sottiaux, is to “bring the power of coding agents to everyone.” Plug-ins connect ChatGPT to tools such as Slack, SharePoint, Airtable and Google Drive. Since OpenAI's developer day on 29 September they have their own place in the sidebar, and OpenAI is adding automations that start when something happens in a connected app.
Codex is the agent for software. Plus subscribers can use it on the web, in the command line, in an editor extension, on iOS and in the cloud, and the free and Go plans get it in the desktop app. In September OpenAI added reusable cloud environments that a team can share with approved settings, a code review view for GitHub and GitLab, and a command line that takes voice. Work and Codex share the same engine and the same permission system.
There were three model releases in a month. GPT-6 Astra arrived on 3 September as OpenAI's most capable model. GPT-6 Sol and Luna followed, and on 29 September GPT-6.1 Sol, which OpenAI says comes close to Astra for agentic coding and professional work at one fifth of the standard token prices.
What is genuinely good
In our view, no assistant offers more breadth. One subscription covers a chat assistant, an agent for documents and a coding agent, with voice, images and a large catalogue of plug-ins. For 20 dollars a month, Plus includes Work and Codex with GPT-6.1 Sol.
Codex suits teams that already work in GitHub. It reads the repository, proposes changes, reviews pull requests and runs tasks in parallel in the cloud while people do other work. Our view from building software with agents every day: the gap between the leading coding agents is small and changes with each release, so the surrounding workflow matters more than the brand.
OpenAI also builds for administrators. In Enterprise and Edu workspaces, plug-ins and the apps behind them are off until an administrator turns them on, and users approve each plug-in's access one by one.
What is not
The tiers are hard to read. There are Free, Go at 8 dollars, Plus at 20, three Pro plans at 100, 200 and 500 dollars, Business at 20 or 25 dollars per user, and Enterprise on request. Limits are given as estimates per five hours. On Plus, OpenAI estimates 5 to 45 local messages with Astra and 350 to 3'000 with Luna, and Work draws on the same allowance as Codex.
Capacity and change can get in the way. On 10 September OpenAI paused new Pro subscriptions because, as Sottiaux put it, “Demand for Astra is really unprecedented.” Models are retired quickly: GPT-5.5 leaves ChatGPT, Work and Codex on 14 October, weeks after its successors arrived. And Astra launched amid questions from researchers about a reasoning technique that can make its thinking harder to monitor, which OpenAI's chief scientist Jakub Pachocki acknowledged: “as model capabilities are increasing, monitorability is getting more challenging.”
The free and Go tiers carry advertising. Ads came to ChatGPT this year and reached India in August. On 5 October OpenAI began adding visual ads next to generated images in the US, with a first group of advertisers. It says they “will also be clearly labeled and won't influence the answers ChatGPT provides.”
Privacy deserves attention. In September OpenAI disclosed that agents in its research environment had posted 53 images to image hosting sites, images that users had uploaded and that had been used in training. It said: “This is not an appropriate use of this data,” and that it could not identify the users to inform them. For companies, OpenAI's own documentation also notes that “Local computer access with Work Cloud does not provide strict zero data retention.”
Lock in and cost
For most teams, Business at 25 dollars per user on monthly billing is the sensible floor: company data stays in a managed workspace, and Work follows that workspace's retention settings. Heavy Codex users will feel the limits and look at Pro or at paying by API key, which leaves out the cloud features.
The lock in sits in habits more than in files. Code written with Codex is ordinary code in your repository, and plug-ins increasingly build on MCP, an open protocol. Chats, shared Spaces and the routines a team builds around them stay with OpenAI.
Our verdict
ChatGPT remains the broadest assistant on the market, and Codex is a serious coding agent for teams that review what it writes. The price of that breadth is a crowded set of tiers, frequent model changes and settings worth checking before company data goes in.
This is how we build Werk, our app platform. AI writes most of the code on a framework we built and trust, and people decide what to build, design how it feels and review every release. The agent can change. The review stays.
