Claude Cowork vs OpenAI Codex: the business case for each

I've been using Claude Cowork as a daily driver since the research preview. Scheduled tasks, skills, vault memory, n8n integrations baked in. It replaced a lot of the glue I used to build by hand. So when OpenAI relaunched Codex on April 16 as "Codex for almost everything," my first reaction wasn't threat assessment. It was curiosity about whether they'd actually solved the same problems differently, or just dressed up the same problems in a new UI.

Having spent time in both: Codex is currently ahead for most knowledge workers. That's the honest take, and I think you're better served by that than by false balance. The gap is specific though, and there are real use cases where Cowork's architecture makes more sense. This post explains both.

What the Codex relaunch actually changed

Before April, Codex was a coding agent. Useful, but narrow. The April 16 update turned it into something closer to a general-purpose workspace: an in-app browser, image generation, persistent memory, local and cloud-hosted scheduled automations, computer use, and a plugin library that crossed 90+ integrations.

The plugin list is the headline for business users. GitHub, Slack, Notion, Google Workspace, GitLab, Atlassian Rovo, CircleCI, and the full Microsoft 365 connector set: Outlook, Teams, SharePoint, and Calendar. These aren't superficial read-only connectors. The M365 connectors support write actions, so Codex can draft and send emails, create SharePoint content, schedule Teams meetings, and act inside the suite rather than just observe it. Codex also supports custom MCP servers, so the ecosystem extends well beyond the default plugin library.

By early April, Codex had crossed 3 million weekly active users with 70% month-over-month growth. By April 21 that figure had hit 4 million. The momentum is real, the UX is familiar to anyone already using ChatGPT, and the integration footprint is the broadest in the category.

The real trade-offs

Context window and compaction

Claude's context window in Cowork is 1M tokens. Codex runs on GPT-5.5, which OpenAI documents as 400K in Codex, though there's a known discrepancy: in live sessions GPT-5.5 in Codex reports an effective context of around 258K, a gap OpenAI has acknowledged in GitHub issue threads but not yet resolved. Codex compensates with aggressive auto-compaction, which works well for most workflows. Long sessions stay coherent because old context gets summarised before it overflows. For most tasks the window size simply doesn't matter. For long-running research, dense codebases, or multi-document analysis where compaction starts losing fidelity, the raw window gap does start to bite.

Computer use: which platforms it actually works on

This is where the framing usually gets reversed. Codex's computer use launched April 16 as a macOS-only feature, and at launch it was also blocked in the European Economic Area, the UK, and Switzerland. Windows support is promised but not shipped. So unless you're a Mac user outside Europe, Codex's headline "operate any app on your computer" capability isn't actually available to you.

Cowork's computer use shipped on both macOS and Windows from the start. Windows Cowork hit general availability on 10 February 2026 with full feature parity, computer use included. For anyone working on Windows, this is a gap in Cowork's favour. That said, a lot of enterprises aren't deploying computer use at all yet. It's expensive relative to standard agent interactions, the error rate in real-world conditions is higher than demos suggest, and many IT teams aren't comfortable with an agent that can freely operate the desktop. So the Windows limitation in Codex may matter less than it looks depending on how much weight you're actually putting on computer use in your evaluation.

Extensibility: skills, plugins, connectors

Both platforms have skills, plugins, and MCP connector support. The underlying architecture is similar even if the terminology differs.

In Cowork, a plugin is a bundle that packages skills, MCP connectors, and sometimes sub-agents into a single installable unit. Skills are the modular instruction sets that shape how the agent reasons about a task. MCP connectors are the transport layer to external systems. Plugins are how you ship those pieces as a single thing someone else can install.

Codex follows the same pattern. Plugins bundle skills and connectors. Custom MCP servers extend the surface further.

The real differences are ecosystem maturity and customisation depth. Codex ships with a much broader default plugin library: 90+ integrations curated by OpenAI, with first-class M365 write actions out of the box. Cowork's library is smaller but its skills model is more open-ended; you can write skills that are pure instruction (no external tools at all) and chain them with vault context and scheduled tasks to build something that feels less like a tool and more like an environment. If you want capable out of the box, Codex wins on ecosystem. If you want to build a personalised operating environment around your own knowledge and rhythms, Cowork's customisation surface is deeper.

The Dispatch gap, and how it closed today

Until today, Dispatch was Cowork's clearest differentiator. Dispatch lets you remote-control a Cowork session running on your desktop from the Claude mobile app. You start tasks, review outputs, redirect work in progress, and monitor sessions running in parallel, all from your phone. It treats Cowork as an async background worker, not a foreground tool you have to sit in front of. (Dispatch is mobile-to-desktop specifically. There's no web Dispatch surface; the phone is the remote and the desktop is the worker.)

Codex had no equivalent. If you weren't at your machine with the app open, you couldn't start or monitor a session.

That changed today. On 14 May 2026, OpenAI shipped Codex control inside the ChatGPT mobile app on iOS and Android, rolling out to all Codex users including the free and Go tiers. Scan a QR code from the desktop app and your phone becomes a control surface for the active Codex session: review outputs, approve commands, switch models, start new threads. Files and credentials stay on your machine; updates stream to your phone in real time. Mobile remote control for Codex on Windows is "to follow." It's not shipping today.

It's not identical to Dispatch. Cowork's mobile interface has had longer to mature for async multi-session work, and the underlying agent in Cowork is set up to run scheduled and background tasks as a first-class behaviour rather than as a session-extension. But the fundamental gap is closing fast. If you'd been staying on Cowork specifically because you wanted mobile control of your AI agent, that calculus shifted today.

Who should use which

My take: Codex has the overall edge right now. Broader plugin ecosystem, native M365 write actions, familiar UX, and a scheduling model that, as of a recent update, can run automations inside an existing thread with context intact or start fresh on a cron. Both automation types can run in the cloud, so they don't depend on your machine being on. For most knowledge workers, that's the right starting point.

But the gap is platform-specific, and two things matter a lot depending on how you work.

Computer use on Windows. Codex's computer use is macOS-only at launch and blocked in the EEA, UK, and Switzerland. Cowork has it on both Mac and Windows, with feature parity since February. For most enterprise knowledge workers on Windows, Codex's headline computer-use feature simply isn't available. That said, see the caveat above: many organisations aren't deploying computer use yet regardless of platform, so weigh this against how central computer use actually is to your evaluation.

Mobile remote control maturity. Codex shipped mobile control today, which is the right direction. But Cowork's Dispatch has been in use longer, and the async-first model, where background tasks run on your desktop while you direct them from your phone, is more established. Whether that maturity gap matters depends on how async your workflow actually is.

The honest breakdown:

• Codex if you're on Mac, outside Europe, and want the broadest integration ecosystem without assembling anything yourself.
• Cowork if you're on Windows and remote control/computer use is central to how you work.
• Either if you're building on custom MCP connectors. Both platforms support the standard, both have plugin ecosystems, and the architecture gap there is smaller than the marketing on either side suggests.

The honest summary

Codex is ahead. It's more capable out of the box, has a wider plugin library, and the UX advantage of the ChatGPT ecosystem matters. For someone starting from scratch today, I'd suggest Codex first.

The cases where I'd still point someone at Cowork: Windows users who need computer use, anyone building a skills-based personalised agent environment, and anyone whose workflow is genuinely async and benefits from a mature mobile remote-control setup. Those are specific conditions, but they describe a lot of enterprise knowledge workers. Know which camp you're in before you decide.

If you're evaluating either platform for your organisation and want an architecture-level view of how it fits into your M365 environment, that's the kind of work I do. Get in touch at consulting@joshwickes.com.