This post began as a follow-up to our spring 2026 Codex introduction. Two practical things matter for use: a clear goal for longer work and the ability to inspect tasks on the move. Current prerequisites differ by surface and account.
The two paths are goal mode and mobile remote control. Goal mode describes the desired outcome and keeps it for the thread. Remote lets supported accounts control a connected computer from the phone. This guide does not promise simultaneous introduction of every function on every plan or surface on one date.
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#
Models, navigation and work surfaces have developed since the original introduction. This article focuses on documented operation: goals in Codex CLI and Remote with a connected computer. Other surfaces may offer more; check the actual account.
The workflow matters more than a fixed list of five apps. The same questions apply everywhere: where does the task run, which material can it access and how will you verify the result?
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.
Goal mode adds persistent goal state to the thread. ChatGPT can also perform multistep work. Choose by which environment reaches the required files and tools and how you can verify the result. Autonomy replaces neither missing sources nor approval for consequential actions.
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 at your desk and check it later on your phone where account and connection support it. Execution location matters: a cloud task can continue while the laptop is closed. A task on your local host needs that computer reachable and the application running. Mobile is remote control; it does not automatically move local work into the cloud.
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.
Current Remote access uses the ChatGPT app with a connected Mac or Windows PC. Availability and permissions depend on rollout and workspace. Check your account’s surface and official setup rather than assuming the earlier launch means identical functions on every plan and mobile app.
Your first Goal Mode moment in five minutes#
The officially documented starting point is Codex CLI version 0.128.0 or later. Open a session in the appropriate environment and set a goal with /goal and your brief. If your web or desktop surface offers a goal feature, check availability and operation in that account. This does not promise a universal web toggle.
For a first test, supply twelve sanitized sample emails as files and two writing samples without confidential data. Brief: “Group these emails into reply now, this week, no reply needed and unclear. Draft replies under 80 words for the first group using the samples. Save the list and drafts; send nothing. Success: every email has exactly one group and unclear cases are marked.” A real inbox later needs supported, authorized access.
Let the task run, then review grouping, open questions and drafts. Style comes from available samples and instructions, not automatically your entire ChatGPT history. Whether you can close a surface depends on execution location: cloud and local hosts have different prerequisites.
The aim is a coherent deliverable with checkable criteria. Compare it with a regular task on the same material. If goal mode adds loops rather than a better result, the simpler path suits that task better.
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.
You do not need to program these tasks yourself. You do need the appropriate environment, required material and authorized access. Simple local file tasks do not require GitHub. Check whether goal mode improves on a regular task by comparing output, effort and errors. Multiple steps alone do not guarantee a benefit.
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.
Long goal tasks consume different allowances depending on model, material and tool calls. There is no reliable universal rule of three or four sessions a week and a fixed Pro multiplier for every task. Check usage, limits and reset time in your account. A bounded goal also helps avoid unnecessary loops.
Choose a plan change based on actual usage and available limits. A longer task alone does not automatically make Pro the right plan.
What is still coming this year#
For new surfaces and integrations, check actual availability in your account. Goal state belongs to the specific thread and does not replace global memory or project documentation. In the CLI, /goal supports setting, pausing, resuming and clearing a goal; other surfaces can differ.
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#
Test goal mode where your account actually offers it. For the documented CLI path, set /goal with a small bounded brief and review the output.
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.
Goal mode can use available project files, native memories or an explicitly connected MCP store. The goal itself does not provide access to external information. First check that the environment can retrieve it and identify the required source in the brief. Our memory guide separates the stores.
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 Setting up an AI system 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.
Sources and product status#
Codex goals and thread state / Codex Remote and the connected computer / OpenAI usage and pricing
