Every few months somebody you trust publishes a careful piece explaining why they have moved from one AI tool to another. They tracked their time. They counted their errors. Nobody paid them to say it. They are usually right about their own work, and that last part is what gets lost by the time it reaches you.
Through 2026 most of those pieces have pointed the same way: leave Claude, move to ChatGPT's Codex app. The reason given is rarely the model. It is that Codex put the browser, the control of your computer and the scheduled jobs in one window, while Claude had them spread across 3 separate places.
4 days after one widely read version of that argument went out, Anthropic shipped the change that answered its main complaint. On 12 August 2026 the Claude side panel in Chrome stopped being a walled-off chat and became the same session as the desktop and mobile apps, carrying your history, your skills and your connectors into the browser with it.
The reviewer was not wrong. The review aged in 4 days, which is what a verdict does when 2 products are leapfrogging each other on a monthly cycle. So the useful thing is not the verdict. It is a way of deciding about your own week.
The 3 capabilities these arguments are always about
Strip the tool names out and every one of these comparisons is about the same 3 things, all of which are new enough that most business owners have never used any of them.
A browser inside the assistant. Rather than you copying text into a chat, the assistant opens pages itself and reads them. Useful for checking competitor pricing, pulling details off a supplier site, or working with any service that has no proper connector. There are 2 versions and the difference matters: a clean sandboxed browser that starts logged out of everything, and a plugin that drives your own browser with you already signed in. The first is for research. The second is the one that can do your admin, and the one that can do damage.
Control of the apps on your computer. The assistant sees the screen, clicks and types in real desktop software. This is what makes the old, stubborn jobs possible, the ones where a program has no API and no export, so a person has always had to sit there and do it by hand.
Jobs that run on a schedule. You write the instructions once and the work happens on a cadence without you opening anything. This is the one that changes how a small business feels, because it turns your assistant from something you visit into something that hands you finished work.
Where the 2 tools stood in August 2026
| Capability | ChatGPT Codex app | Claude |
|---|---|---|
| Browser inside the app | Built-in browser, April 2026 | Built-in browser in the desktop app, July 2026, sandboxed and logged out by default |
| Drives your own logged-in browser | Chrome plugin | Claude in Chrome, and from 12 August 2026 the side panel runs the same session as the apps |
| Clicks and types in desktop apps | macOS April 2026, Windows May 2026, via the Computer Use plugin | Computer use in Cowork, a research preview on Pro and Max |
| Scheduled jobs | Generally available March 2026, plus triggers from Gmail, Slack and GitHub from 25 August 2026 | On every paid plan, and they run in the cloud, so a shut laptop does not stop them |
| Makes images in the chat | Yes, using ChatGPT's image model | No. Claude builds charts, diagrams and design files from code, and reaches image models through a connector |
| Moves your setup in | Settings, then Import: Claude Code, Cowork and Cursor, from 11 August 2026 | Not applicable |
That table is a snapshot and it is the part of this guide with the shortest life. Check it yourself before you spend anything on the strength of it.
The difference that will still be true next year
Feature parity keeps arriving. What does not change is where the work physically happens, and for a small business that is the question worth asking.
Anything that drives your own logged-in browser or clicks around your own desktop needs your computer awake, unlocked and left alone while it works. That is true of both tools, because it has to be. Anything that only needs the web, your files in the cloud and your connected accounts can run on somebody else's machine while your laptop is shut.
2 of our own scheduled jobs sit on the cloud side of that line. One runs at 2pm every Tuesday, researches what moved in AI that week and writes 4 ready-to-record scripts into a folder for review that evening. Another runs at 4.30pm on a Thursday, finds one podcast episode or video worth listening to, checks it is real and the right length, and emails it. Neither needs a machine to be on, so neither fails on a busy day.
The job that would sit on the other side of the line: logging into a supplier portal that has no API, reading yesterday's orders off the screen, and putting them somewhere useful. That needs a real signed-in browser session, so it needs your computer running at 6am and nobody touching it.
Work out which of your 2 or 3 recurring jobs is which. That answer survives every update either company ships.
The 2-week test
The mistake is switching on the strength of someone else's 5 jobs. The fix costs you 2 weeks and one month of paying for both.
Pick 5 jobs from your real week. Not test questions, and not the impressive thing you would demo to someone. Use the ones you repeat: the customer email you rewrite, the numbers you summarise, the thing you look up on 3 sites, the post you draft, the admin task you keep putting off.
Give both tools the same start. Most comparisons are unfair because the old tool has months of context and the new one has none, so paste this into each of them before you begin.
I am testing you against another AI tool for 2 weeks, on the same jobs, so we are starting level.
My business: [what you sell, to whom, where]. My week: [how many hours you have and when]. What I care about in an answer: [for example, that it sounds like me, that numbers are right, that it is finished rather than a draft].
Do not give me options unless I ask for them. If something is missing, ask one question and then get on with it. When you finish a job, tell me in one line what you were unsure about.
Then keep a log. 5 rows per tool, filled in as you go rather than from memory at the end.
| Job | Minutes it took | Corrections you had to make | Finished, or handed back to you | Apps you had to open yourself |
|---|
The decision rule: switch only if the new tool wins on 3 of the 5, and only if those 3 are jobs you do every week. A tool that is better at your rarest job is not better at your work. If it comes out level, stay where you are, because the setup you already have is worth something the scorecard does not show.
What a switch costs before you count the subscription
Your saved instructions, projects, files and skills are the reason it hurts to move. OpenAI added an import in August 2026 that covers Claude Code, Cowork and Cursor, and it will keep the imported work in sync afterwards, so the file part is largely solved. Your own habits are not imported. Expect a fortnight of being slower.
Pay for both during the test and cancel nothing on day one. A month of overlap is the cheapest part of this decision, and cancelling first is how people end up committed to a tool they have not really tried.
Where this guide stops
We have not moved our own business onto Codex, so what you have here is the test rather than a result. Anything a review tells you about which tool is smarter is a claim about their 5 jobs, ours included.
One thing worth checking early, because it is the kind of thing no review will predict for you: if part of your work sits near a sensitive area, hiring and recruitment being the common one, either tool may decline the job while the other does it without comment. Run your one awkward job in week one, not week 2.
And when you do turn on a browser plugin or let an assistant use your computer, remember what it is holding. Signed in as you, in your browser, it can send, buy, post and delete as you. Approve those steps yourself, keep the unattended scheduled jobs to work that only reads and drafts, and give it the awkward permission the day you have time to watch it, not the day you are late.
Run your 5 jobs. The answer will be about your week rather than anyone else's.