OpenAI is positioning its Codex tool as a productivity app, Anthropic has Cowork, and a viewer DM’d me asking the obvious question: which one would I actually pay for? After a couple of weeks living in both, I finally put them head-to-head on a real piece of my work. The answer isn’t really about features — it’s about two completely different models of how AI works for you.

Two Different Models of AI Work

The first model is autonomous productivity: you hand the AI a task, it executes — reading files, running processes, producing output. That’s Codex, and it’s powerful for discrete, contained, handoff-friendly work. The second model is embedded productivity: the AI is woven into your workflow. It reads your past emails, knows your clients, and doesn’t just do the task — it does it as someone who knows you. That’s Claude Cowork in a nutshell. Almost every difference I ran into traces back to this split.

The Test

I gave both tools the same briefing against a real client folder: a project invoice that needed to match my logged hours, a spreadsheet tracking burn rate, and three follow-up emails to draft — one for the client, one for my VP, and one for my team. Codex plugs into your computer and reviews the materials directly; with Cowork, you connect it to a specific folder. Same inputs, same ask, two very different outputs.

How Codex Did

Codex was fast and, for the most part, correct. It produced the deliverables I asked for and showed its reasoning. If the job had been “find the discrepancies in these files,” it would have nailed it. But the output was basic, and the emails gave it away — generic. They covered the right project and the right information, but they didn’t sound like me. Codex had the task; it just didn’t have the context. That’s autonomous productivity doing exactly what it’s built for, and no more.

How Cowork Did

Cowork is built for exactly this kind of work. It produced the project tracker with more complete formulas and a cleaner layout, drafted all three emails, and wrote the budget the way I actually write budgets — using the templates I’d given it, not a generic format. Because it could pull context from the projects connected to that folder, the output read like it came from someone who knows my work. The natural question is whether this is just a context-access problem Codex will patch in the next update. I don’t think so — it’s a difference in model, not a missing feature.

So Which Should You Pay For?

Here’s the honest answer. If your work is task-shaped — defined goals, contained outputs, things you can brief clearly and evaluate cleanly — Codex is genuinely good and worth paying for. If your work runs on context, like client relationships, ongoing projects, and communication that needs to sound like you and draw on your history, Cowork is the better buy. Most people whose work lives in documents, emails, and client relationships fall in that second bucket, which is why I’d point most folks toward Claude. The question was never which one is smarter — it’s which model fits the shape of your work.

Whichever one matches your job description is the one worth paying for. If either stood out to you, drop a comment — I’m curious which way people lean. And if you want to see how Cowork works inside Microsoft, check out the related video above.