Claude Cowork vs ChatGPT Work: Agent Modes Tested

> **TL;DR:** Claude and ChatGPT have converged on the same three-part product: a chat mode, an agent mode that does the work (Claude Cowork / ChatGPT Work), and a build mode (Claude Code / Codex). In hands-on comparison the agents behave differently — ChatGPT tends to return one finished deliverable, while Claude browses live, asks clarifying questions, saves real files and itemizes its uncertainty. The biggest practical split is where tasks run: ChatGPT's scheduled tasks run on their own, Claude's only run while Cowork is open on a machine that's powered on.

Key Takeaways

- Both apps now ship the same layout: chat, an agent mode, a build mode, and Projects in the sidebar. - ChatGPT Work shipped quietly and reached 10 million users in under two weeks; Claude Cowork is its direct counterpart. - Claude's scheduled tasks only run while Cowork is open — ChatGPT's run without the machine being on. - Record a Skill replaces prompting with demonstrating: one narrated workflow became a 108-step skill in Cowork. - ChatGPT organises plugins as a searchable one-click library; Claude ships pre-built bundles (its data plugin arrives with ~10 skills and 8 connectors).

Open ChatGPT and Claude side by side right now and you will struggle to tell the navigation apart. Both have landed on the same three-part structure: a **chat mode** for asking, an **agent mode** for doing, and a **build mode** for shipping. On Claude that reads Chat / Cowork / Claude Code. On ChatGPT it reads chat / Work / Codex. Both keep Projects in the sidebar to hold everything for one job in one place, and on that feature the two are effectively equivalent.

The convergence is the story. Two labs with different research cultures independently decided that a chat window is only a third of a product — and the interesting part is what happens once you actually put the agents to work on the same brief.

The three modes, decoded

Chat is now the smallest piece

Chat is still the default surface and still the thing most people mean when they say "I use ChatGPT." It is also, on both platforms, the part that has changed least. The competitive action has moved one tab over.

The agent mode is where the work happens

**Claude Cowork** is toggled at the top of the Claude interface, right next to Chat. Rather than answering about a task, it operates on your own computer and your own apps to complete the whole thing — Anthropic positions it as the no-code place where real work gets done.

**ChatGPT Work** is the functional counterpart: an agent mode inside ChatGPT for creating, learning and exploring. It shipped quietly and reached 10 million users in under two weeks, which is a useful signal about how much latent demand there was for "do the task" rather than "describe the task."

The build side is not developer-only

**Codex** is ChatGPT's mode for building, debugging and shipping apps, websites and tools. It gets misread as a developer product, and that misreading costs non-coders the single most capable surface in the app. **Claude Code** is the equivalent build-and-ship side on Claude, sitting alongside chat and Cowork. If you have never opened either because you assumed you needed to write code, that assumption is the thing to drop first. More coverage of these surfaces lives in our [New AI Tools & Skills](https://speka.info/new-ai-tools/) hub.


The trip-planning test: same brief, two work styles

We gave both agents an identical, deliberately messy brief — a PDF of scattered preferences and a $2,000 budget — and let them run.

**ChatGPT** returned a complete package inside the thread: a full itinerary, a booking checklist with live links, and a budget spreadsheet landing at $1,681.

**Claude** worked differently. It browsed live, including Google Flights, paused mid-task to ask which dates to build the trip around, explained *why* it picked the neighbourhood it did, and saved three real files to Google Drive — among them an $1,800 budget sheet.

The pattern that emerges holds across most of the tests below. ChatGPT optimises for a finished deliverable handed back to you in one motion. Claude optimises for artifacts that exist outside the chat and a visible trail of its reasoning, at the cost of interrupting you to ask. Neither is strictly better; they suit different tolerances for being asked questions.

Scheduled tasks: same feature, different leash

Both platforms can now run tasks on a schedule, and this is where the sharpest practical difference sits.

**ChatGPT** offers three presets — a daily brief, a weekly review, and a follow-up monitor. We set up a Slack-sourced weekly review pinned to Fridays, and it ran instantly on demand without asking for extra setup.

**Claude Cowork** has the same capability but flagged two limits up front, to its credit: it can only run while Cowork is open, which means the machine has to be on; and with mail and calendar unconnected, it could only draw from Slack.

That is the cloud-versus-computer split in one sentence. Agent tasks that execute in a provider's cloud keep running while your laptop is shut; agent tasks that execute locally do not, but they also never hand your working files to someone else's infrastructure. Which trade you prefer depends on how you feel about where model execution actually happens — a question that gets less abstract once you read [what an OpenAI model did when tested for sandbox escape](https://speka.info/blog/openai-model-sandbox-escape-what-the-test-showed).

Plugins, skills and connectors: a library versus a bundle

**ChatGPT** exposes a plugins section that merges apps, MCPs and skills into one browsable library, sorted into categories — productivity, creativity, developer tools, business and operations, data and analytics, communication, finance, healthcare, travel. Each entry is a one-click install-and-connect.

**Claude** takes the opposite approach. Connectors are added through Settings → Connectors, and its plugins arrive as ready-made bundles built around a use case rather than individual components. The data plugin, for example, ships with roughly 10 skills and 8 connectors already installed, so you are not assembling a stack one checkbox at a time.

Library versus bundle is a genuine philosophical difference: ChatGPT assumes you know what you want, Claude assumes you want the job pre-wired. Both let you upload your own skills and connectors, pull from a marketplace, or build new ones from scratch.

Record a Skill: demonstrating instead of prompting

The most underrated feature on either platform requires no prompting at all.

On **ChatGPT**, Record a Skill lives under Create. It records you performing a task once and converts that recording into a reusable skill the assistant can rerun. No code involved.

On **Claude Cowork**, the same capability was added recently. We recorded a weekly workflow while narrating it aloud, and Cowork turned it into **108 discrete steps** that can be invoked with a single command afterwards.

That is a meaningful shift in how you instruct these systems. Instead of writing an increasingly baroque prompt describing a process, you do the process once, out loud, and hand over the recording.


The Slack test: one number versus a full ledger

We pointed both agents at a real work Slack and asked how many content releases had gone live since June 1.

**ChatGPT** returned a single deduplicated figure: 107 releases. Clean, quotable, done.

**Claude** returned the full itemised list broken out by month — 9 in July, 13 in June — counted short-form pieces separately, and explicitly flagged two entries it was not certain about.

Again: one answer versus one audit trail. If you are pasting a number into a deck, ChatGPT's output is what you want. If you are going to be asked "where did that number come from," Claude's is.

Where they genuinely diverge

Not everything has a mirror on the other side. **ChatGPT Sites** lets you build and publish a live website for free, and there is no direct Claude equivalent in this comparison. Claude counters with **Artifacts and Live Artifacts**, which produce interactive, runnable outputs inside the app rather than a public URL.

Slash commands are worth knowing on both. `/side` lets you ask a question while an agent task is already running without derailing it — small feature, disproportionate quality-of-life gain when a long task is mid-flight. `/pet` builds your own AI pet, which is the entry point most people will actually use to understand what a custom skill *is*; free pixel-pet sprite libraries pair with it nicely.

What this actually means

The strategic read is that neither company thinks the chat box is the product any more. Both are betting that the durable surface is an agent with access to your files, your Slack, your calendar and your schedule — and that the moat is integrations, not model quality alone.

That framing matters because model quality is commoditising underneath them. Open-weight releases like [Kimi K3 landing on Hugging Face](https://speka.info/blog/kimi-k3-open-weights-land-on-hugging-face) keep compressing the gap on raw capability, while policy pressure of the kind we covered in [the Chinese AI model ban analysis](https://speka.info/blog/chinese-ai-model-ban-what-it-would-cost-us-firms) reshapes which models firms are even allowed to run. Interfaces, connectors and recorded skills are stickier than weights.

Practical advice: if you want tasks that run on a schedule whether or not your laptop is open, start with ChatGPT Work. If you want an agent that touches real files on your own machine and shows its work, start with Claude Cowork. If you have been avoiding Codex or Claude Code because you don't write code — open them anyway. That is where the ceiling is.

Frequently Asked Questions

What is Claude Cowork?

Claude Cowork is an agent mode inside Claude, toggled at the top of the interface next to Chat. It operates on your own computer and connected apps to complete entire tasks rather than just answering questions about them.

What is ChatGPT Work and how is it different from Cowork?

ChatGPT Work is OpenAI's agent mode for creating, learning and exploring — the functional counterpart to Claude Cowork. It shipped quietly and reached 10 million users in under two weeks. The main practical difference is that ChatGPT's scheduled tasks run without your machine being on, while Cowork tasks only run while Cowork is open.

Do you need to code to use Codex or Claude Code?

No. Both are build-and-ship modes for apps, websites and tools, and both are commonly misread as developer-only. Non-coders can build in either.

What does Record a Skill do?

It records you performing a task once and converts the recording into a reusable skill the assistant can rerun on command. In one Cowork test, a narrated weekly workflow became a 108-step skill invoked by a single command, with no coding involved.

How do plugins and connectors differ between the two apps?

ChatGPT offers a browsable library of apps, MCPs and skills sorted by category with one-click install. Claude adds connectors via Settings → Connectors and ships plugins as pre-built bundles — its data plugin arrives with roughly 10 skills and 8 connectors already installed.

Which agent is better for research-heavy tasks?

Claude tends to browse live, pause to ask clarifying questions, explain its choices and save real files, while ChatGPT tends to return a single finished deliverable in the thread. Pick Claude when you need an audit trail, ChatGPT when you need a clean answer fast.