In mid-April, OpenAI Codex had two million weekly users. Six weeks later it has four million. The growth was not the news. The news was what shipped in between to make the next two million possible.
Two updates in particular matter for people who already use ChatGPT and read our April Codex guide. The first is Codex Mobile, which left the Plus-only bucket on May 14 and is now part of every ChatGPT plan including Free and Go. The second is Goal Mode, which exited the experimental flag on May 21 and is now generally available across the web app, the desktop app, the IDE extension and the CLI.
Neither of these is a new tool. Both are upgrades that change how Codex fits into a real workday. This article is for readers who liked the April beginner's guide and want to know what the May updates actually do, without the OpenAI marketing copy and without code.
What changed in six weeks
Six features shipped between April 16 and May 22. The user count doubled. GPT-5.5 became the default model behind Codex. Codex Mobile got freed up to every plan. Goal Mode reached general availability. Appshots, which let you attach any macOS window to a Codex thread by hotkey, also went general. A Chrome extension launched that lets Codex work inside live browser sessions instead of in a sandbox.
That is a lot for six weeks. The April guide covered five surfaces. The list has not grown, it has deepened. The same five Codex environments (Web, iOS, Desktop, VS Code, CLI) now do more with less setup. Two of those changes are big enough to be worth their own walkthrough.
Goal Mode in one sentence
Goal Mode is the difference between asking Codex to do a task and asking Codex to reach an outcome.
In normal Codex use, you give it an instruction like "draft this email" or "summarize this PDF." Codex returns a result and waits. If the result is not quite right, you give it follow-up instructions. The interaction is turn by turn.
In Goal Mode, you describe the end state and the success criteria. Something like "prepare a clean handoff document for the new freelancer covering everything she needs for the next two weeks. It should be self-contained, written in plain English, and include links to the three relevant Notion pages." Codex breaks that into steps on its own, executes them in sequence, checks its own work against your criteria, and only comes back when it thinks it is done or when it needs your decision on something it cannot resolve.
This is what the April post promised in one line that turned out to be the most quoted line of the whole article. "Codex executes, ChatGPT replies." Goal Mode is the version of that promise that is actually delegation rather than execution. The previous Codex felt like an assistant who needed instructions for every step. Goal Mode feels like an assistant who reads the brief and gets back to you when there is a question.
What Codex Mobile actually is in May 2026
Mobile is the part most people get wrong. Codex Mobile is not a smaller Codex that runs on your phone. It is a remote control for the Codex that already runs on your Mac.
You start a task on the desktop. You walk to a meeting. The phone shows you the live progress, lets you read intermediate outputs, approve actions Codex wants to take, switch models if the task is heavier than expected, or start a new task that runs in parallel. When you come back to the desk, everything is in the same state. The model did not pause when you closed the laptop. It kept working because the work was never on your laptop in the first place.
Two practical consequences come from this. First, you can decide that the Mac stays plugged in and unlocked at the office or at home, and the phone is your travel device. The Mac is the workshop, the phone is the dashboard. Second, you stop thinking of mobile as a downgraded version of the real tool. Mobile is the part that goes with you, not the lesser version of what stays at the desk.
The May 14 launch removed the Plus-paywall on this. Everyone with a ChatGPT account, including Free users and ChatGPT Go subscribers, can install the latest ChatGPT app on iOS or Android and find Codex inside it. The Mac host requirement still applies for the full feature set, but the basic monitoring and task-handoff functions work on phone alone.
Your first Goal Mode moment in five minutes
Open Codex on the web at chatgpt.com/codex. Look for the toggle labeled Goal Mode in the input area, or use the slash command /goal. Both routes work.
For the first try, do not pick a coding task. Try something like this. "Goal: I have an unread inbox with twelve emails from clients. Read all of them and prepare a triage list. Group them into reply now, reply this week, do not need a reply, and unclear. For each one in the reply now group, draft a short response based on what I usually write. Success criterion: I can act on the triage list in under ten minutes."
Hit run. Codex will work for several minutes, sometimes longer than expected. You can close the browser. When it is done, you have a structured triage list and three or four draft replies that match your own tone if you have been using ChatGPT consistently long enough that the model knows how you write.
The first time this happens, the difference from regular ChatGPT is obvious. ChatGPT would have replied to one email at a time. Codex in Goal Mode worked through all twelve, made its own classification, drafted responses, and packaged everything into a single deliverable. That is the move you do not get from a chat interface.
Ten Goal Mode tasks that have nothing to do with code
These are the tasks where I have seen Goal Mode deliver consistent results for solo operators and small teams in the last two weeks.
Triage a client inbox at the start of the day and produce a sorted action list. Take a meeting recording transcript and produce both meeting notes and a to-do list for each participant. Read a long PDF contract and produce a list of every change since the previous version. Translate a multi-page document and run a glossary consistency check so the same term is rendered the same way every time. Take a folder of receipts as images and produce an Excel-ready table with date, vendor, amount and category. Compare two Notion pages and produce a third page that merges them without duplicates. Take an outline and turn it into a slide deck with speaker notes. Prepare a freelancer handoff covering background, current state, deadlines and contact list. Read fifty job applications against a job description and produce a ranked shortlist with one-line reasoning per candidate. Run a competitive scan across five named competitor sites and produce a one-page summary of what is new.
None of these require GitHub, code, or technical setup. All of them benefit from Goal Mode because the work is not one prompt and one reply. The work is a sequence of small judgments that adds up to an outcome.
What Codex Mobile is for in the workday
The mistake I see in early Codex Mobile use is starting big tasks from the phone. That works, but it is not the strength. The strength is monitoring and handoff.
In practice, the pattern that works looks like this. At the start of the workday, on the desktop, you set up two or three Goal Mode tasks that will take twenty to forty minutes each. You go to a meeting, a workout or a coffee. The phone shows you live updates. When a task asks for a decision, you tap to approve or redirect. When a task finishes, you see the result and either accept it or ask for changes. By the time you are back at the desk, two of three tasks are complete and one is waiting for your input.
The other Mobile pattern is starting small tasks while you are out. Something you would have written down in Notes to do later becomes a Codex task started immediately from the phone. By the time you get to the desk, the first draft is already there.
The honest limit is that Mobile is dependent on the Mac being awake and online. If your laptop is closed in your bag, the cloud version of Codex still runs but the Mac-host features pause. For pure cloud tasks, Mobile works the same as Web. For tasks that need your local environment, the Mac has to be available.
Goal Mode versus normal Codex tasks
Goal Mode is not always the right choice. Normal Codex tasks still make sense for fast one-shots where you know exactly what you want and the work is a single transformation.
Use a normal task when you want a single output and you can describe it precisely in one sentence. "Rewrite this paragraph to sound less formal." "Translate these three sentences to Spanish." "Format this list as a Markdown table." These are not goal-oriented. They are transformations. Codex handles them faster without Goal Mode.
Use Goal Mode when the work is multi-step, when success depends on a sequence of decisions, when you want to specify outcomes rather than steps. The triage, the handoff document, the contract diff, the competitive scan. These benefit from the model planning its own approach instead of executing your micro-instructions.
A useful heuristic. If you would explain the task to a junior assistant in one breath, use normal Codex. If you would brief a freelancer for a half-day project, use Goal Mode.
The stumbling blocks I have seen
Three patterns trip people up in the first two weeks of using Goal Mode.
The first is over-defining the success criteria. "Success: the document is clear, complete, accurate, well-formatted, free of errors, and matches my style" reads like a thorough brief but tells Codex nothing it can check against. Codex evaluates its own work against criteria it can measure. "Success: every email has a draft reply under 80 words" is checkable. "Success: the document is good" is not.
The second is treating Goal Mode like a fire-and-forget machine for sensitive work. Goal Mode is autonomous in its planning, not in its judgment. A finished output still needs your read. The point of Goal Mode is that you read once at the end instead of reviewing every step, not that you never read.
The third is forgetting that Codex still has limits on long-running tasks per plan. Plus users will hit the wall on three or four long Goal Mode sessions per week. Pro users have ten times that. The cost calculation for which plan to pick changes the moment you start doing real work in Goal Mode, because each session uses more compute than a regular task.
If you bumped from Plus to Pro after the April guide, you are in the right plan. If you stayed on Plus and Goal Mode now fits your workflow, the math has shifted.
What is still coming this year
The same product strategy that surfaced in March is still on the rails. OpenAI is collapsing ChatGPT, Codex and the Atlas browser into a single desktop super-app over the rest of 2026. The April update was one step. May was another. The next steps that have been hinted at publicly include persistent cross-app memory, deeper Goal Mode integration with the Atlas browser, and a unified billing model that stops separating the chat side from the agent side.
If you spent six weeks learning Codex, none of that is wasted. The concepts are the same. The packaging gets simpler.
When Goal Mode is not for you
If you mostly use ChatGPT to look up facts, draft single short messages, or have conversations to think out loud, Goal Mode is overkill. The chat surface does that work better and faster.
If you do not have at least one recurring task per week that takes longer than fifteen minutes to do yourself, Goal Mode will not give you back enough time to justify the learning curve.
If your work is highly sensitive and you cannot have any task run autonomously without a human in the loop on each step, normal Codex with manual approvals is closer to the right tool.
For everyone else, the math is straightforward. Pick two recurring tasks. Convert them to Goal Mode briefs once. Reuse those briefs every week. The time saving compounds because you stop writing the same instructions over and over.
What to do next
If you read the April guide and never went past Codex Web, go back to chatgpt.com/codex and switch on Goal Mode for the next task you would normally type as a chat. Watch what changes.
If you read the April guide and have been using Codex Web regularly, install the ChatGPT app on your phone if it is not there already, and let Mobile run as a dashboard while you set up a Goal Mode task at the desk. The first time the phone tells you the task is done while you are in a meeting, the pattern clicks.
If you read our follow-up on Codex memory and connected an MCP memory layer, Goal Mode reads from the same memory. A Goal Mode brief can reference past decisions, client profiles and project notes without you pasting them in each time. That is when the workflow stops feeling like AI and starts feeling like a team.
And if you want help wiring Codex, Goal Mode, mobile and a memory layer into the actual workday for a small team, that is what we do at StudioMeyer. The KI-Mitarbeiter setup is a six-week build that does exactly this for SMEs who have been paying for ChatGPT Plus for months and never quite turned it into a working tool.
