ChatGPT Work vs Claude Cowork: the answer is in your admin console
Two workspace agents shipped 48 hours apart. The thing that actually decides which one your team runs isn't the model — it's a toggle an admin flipped before anyone got to try either. Here's the seam, and my routing rule.
- #chatgpt
- #claude
- #agents
- #workflows
Anthropic put Cowork on web and mobile on 7 July 2026. OpenAI announced ChatGPT Work on 9 July. Forty-eight hours, two products that describe themselves almost identically: hand over an outcome, the agent works across your apps for hours, you get finished artifacts back.
So every comparison online is a feature table. Feature tables are useless here, because as an EM picking tooling for a team I discovered the decision was mostly already made for me — by plan defaults I hadn't read.
The default nobody mentions
Cowork is on by default for Team and Enterprise organizations. But the second toggle, "Run Cowork in the cloud", ships enabled on Team and disabled on Enterprise — on Enterprise an owner has to switch it on and then grant the capability to specific groups through custom roles. Same product, opposite starting position, decided by which contract finance signed.
ChatGPT Work went the other way: available on Pro, Enterprise and Edu the day it was announced, with Plus and Business following within days. The desktop app is global, on Mac and Windows, on every plan including Free.
Read that again. On Enterprise, Anthropic's agent starts locked and OpenAI's starts open. Your team's "which is better" debate is downstream of that.
What each one actually reaches
This is the honest seam, and it isn't cloud-versus-local — both do both now.
ChatGPT Work reaches your SaaS. A unified plugins directory covers Slack,
Microsoft Teams, Google Drive, SharePoint, Gmail, Outlook, Salesforce, Jira,
Zapier and Adobe Acrobat, referenced inline with @. It runs on GPT-5.6, ships
scheduled tasks, a built-in browser, Computer Use for background execution, and
Sites (public beta) for live dashboards. It inherits Codex's governance model —
5 million weekly Codex users, over a million of them outside software.
Cowork reaches the folder. Desktop is still where the local files and the browser live, and Anthropic is explicit that a project tied to a local folder supports Cowork sessions on desktop only. Cloud sessions are the ones that survive with no device online and run the 6am scheduled task.
The part that decides it for a team
Not capability — evidence. Cowork now emits OpenTelemetry events for tool calls, connector invocations, file operations, skills used and approval status, into Splunk or Cribl. Admins get per-tool connector controls, so you can leave a connector readable and switch its writes off org-wide. Enterprise adds SCIM groups with custom roles and group-level spend limits.
ChatGPT Work answers the same question differently: spend limits at workspace, group and individual level in the Admin Console, network access policies, a Compliance API, and an auto-review step where a stronger model inspects important API-touching actions before they run.
The difference is where the check sits. Anthropic makes the human approval a field you can query later. OpenAI makes a model the reviewer so the human doesn't have to be in the loop.
My routing rule
I don't route by task type. I route by whether the write is reversible.
- Reversible cloud writes — a draft deck, a Jira comment, a research doc, a dashboard. ChatGPT Work. Momentum beats a checkpoint when the undo is one click, and its reach into the SaaS your company already pays for is wider.
- Irreversible or local writes — rewriting a folder, a spreadsheet of record, a connector that can write into Salesforce. Cowork, on desktop, because "nothing ships until you've reviewed and approved it" is the product's stated contract and the approval lands in your telemetry.
- Both, deliberately. Twelve months from now the ranking will have flipped at least twice. Pay for the one your compliance story needs and keep the other on a personal plan so switching costs you a session, not a quarter.
Do this today
Open both admin consoles and write down three values: whether "Run Cowork in the cloud" is on, whether the skip-approval setting for write-capable connectors is on, and what your ChatGPT workspace default spend limit is. Those three lines are your team's real AI policy. Whatever the strategy deck says.
Sources
Keep reading
More guides like this one.
A 3-Step AI Cost Audit: Seats, Tokens, and the Hours Nobody Books
Two thirds of companies have an AI dashboard. Barely a quarter can actually see what AI costs them. Here is the 3-step audit I run on a tooling budget: what to list, which number to pull for each line, and the threshold that means kill it.
Audit Your Own Job for AI (the Prompt I Run on My Role)
AI won't take your whole job — it takes tasks inside it. Here's the audit prompt I paste into Claude to score every weekly task for delegation, plus the EM rule I use to decide what stays human.
My No-Overwhelm System for Staying Current on AI
You cannot read every AI newsletter and still ship work. Here is the small, fixed stack I subscribe to — one pick per job, across newsletters, video, creators, communities, and podcasts — plus the 15-minute weekly habit I use to triage it.