Work From Anywhere Just Got Tested by Two AI Labs, 48 Hours Apart

I ran it from 6,000 miles away for 20 days straight. This week, two AI labs made it official.

TLDR: On July 7, Anthropic put Claude Cowork on mobile and web, so a background task can run while your laptop is closed. On July 9, OpenAI answered by merging Codex into ChatGPT and launching ChatGPT Work, an agent that turns a goal into a finished spreadsheet, deck, doc, or website instead of a chat reply. Both are usage-metered, not flat-fee. Pick one workflow you already know cold, run it with approvals switched on, and measure it against your actual hours before you schedule anything recurring.

“Work from anywhere” has been a slogan for a decade. This week it became a testable claim. This newsletter got written by an agent running inside Claude Cowork, on a schedule, before I was at my desk. That’s not a metaphor. I’ve been running Cowork through Dispatch as a daily tool for about two and a half, three months now, and it’s the literal infrastructure behind the issue you’re reading right now. So when Anthropic pushed Cowork out to phones and browsers this week, I wasn’t reading a press release about someone else’s tool. I was reading about the thing that already runs part of my business.

Then two days later, OpenAI answered.

By the end of this you’ll know what each launch actually ships, not the marketing description. You’ll know which one to pilot first depending on how technical your team is, and the one guardrail I’d never turn off no matter which agent you pick.

A few things worth flagging up top:

• Neither of these is a better chatbot. Both are agents that take a goal and work for hours without you watching. Both hand back a finished artifact: a sheet, a deck, a doc, a live web page.

• Both bill by usage, not by seat. Nobody published per-task prices. Budget the first week like an experiment, not a subscription.

• The tester numbers in both launches come from the vendors. Real names, real workflows, real self-reported results. Still the vendor’s own case studies. Treat them as hypotheses, not proof.

Who this is for: any operator running a small team who already delegates one recurring, multi-step task to a person. A report, a competitor scan, a monthly deck. You’re deciding whether to hand the first draft of that task to an agent instead.

The 48 hours that mattered

Anthropic’s move landed first. Claude Cowork launched as a desktop app back in January, built to feel like Claude Code without the code. On July 7, Anthropic expanded it to web and mobile for Max subscribers. The pitch is simple: start a task from your desk, get a status update on your phone, and come back to a finished output even if your laptop is closed the whole time. Anthropic’s own example: “Set Monday’s client prep for 6am: Claude works through the email threads, transcripts, and recent news, builds the briefing doc, and leaves the follow-up email drafted but unsent. Review it over coffee.”

I can vouch for the part about the laptop staying closed. I ran Dispatch on my phone for 20 straight days this summer on a trip to Switzerland, with the host device where the actual work was running sitting 6,000 miles away back home. Zero issues the entire trip. That’s the pitch Anthropic is making to everyone else this week. I’d already tested it at the most inconvenient distance I could put between me and my desk.

Two days later, on July 9, OpenAI answered with ChatGPT Work, an agent built on the newly released GPT-5.6 with Codex technology built in. Same week, OpenAI merged the standalone Codex app into one ChatGPT desktop app. Chat, Work, and Codex now live in one place, on every plan including Free. OpenAI also announced it’s sunsetting the standalone Atlas browser, folding those agentic-browsing features into ChatGPT itself.

Two labs, the same week, converging on the same bet: the fight isn’t about who has the sharper chatbot anymore. It’s about who owns the place where the actual work gets done.

What “ships finished work” actually means

Regular ChatGPT and regular Claude answer a question in the moment. Both of these products do something different. You give them an outcome, not a prompt, and they gather context from your connected tools, plan the steps, and keep working, sometimes for hours, without you sitting there.

Cowork’s early data backs up what that looks like in practice. Anthropic sampled 1.2 million anonymized Cowork sessions across more than 600,000 organizations over two weeks in May. The biggest category, at 33.4%, wasn’t coding. It was what Anthropic calls “business process operating”: pulling scattered updates into one report, building onboarding checklists, reconciling spreadsheets, the stuff that’s part of a lot of jobs but rarely anyone’s actual title. Content creation and copywriting came in second at 16.4%. Software development, the thing Cowork’s ancestor was built for, was 8.7%.

ChatGPT Work is chasing the same territory from the coding side. OpenAI says Codex already had 5 million weekly users and, more tellingly, more than 1 million people were using it for work that had nothing to do with software before Work even had a name. The new agent connects to more than 1,400 plugins and runs in Plan mode, proposing a step-by-step plan you approve before it touches anything. Its output format of choice is now “Sites,” a live web page instead of a static document. Sites update themselves as the underlying data changes.

Real numbers from real early testers

The launch materials name real people at real companies, which is more than most product announcements bother with. Worth reading with one caveat attached: these are vendor-arranged case studies, not independent audits.

Angela Ferrante, Head of Enterprise Marketing at Zapier, used ChatGPT Work to trace lead journeys across HubSpot, Gong, and email. Before, inspecting a single lead properly took her team 35 to 45 minutes, which means most leads never got that inspection. She put the result plainly: “That helped us identify and hand off seven figures in pipeline every month to sales.”

Nathan Bolt, Head of Digital Products at Virgin Atlantic, ran competitor customer-journey benchmarking that used to take his team weeks per cycle. His read: “A competitor analysis cycle that would normally take weeks now takes hours, helping us move from insight to product decisions much faster.”

On the Cowork side, one small business owner described it this way in an independent review. “I’ve been using Claude Cowork daily, and this is the first time an AI tool has genuinely changed how my workday flows,” they wrote. “Task queueing is the real unlock.”

Walk away with: three real, named examples of the exact workflow class both products are best at right now: recurring reporting, lead or customer-journey triage, and the monthly deck nobody wants to build by hand.

Pick your pilot workflow

The move: point one of these agents at one recurring task you already know intimately, not a reorganization of how your team works.

Here’s the filter I’d run before picking a first workflow:

FIRST AGENT PILOT, PICK ONE

1. Do you already know the current hours-per-cycle for this task?
   If you can't name the number today, you won't be able to
   measure whether the agent actually helped. Pick a task you
   can time first.

2. Is the output a document, sheet, deck, or web page,
   not an irreversible action?
   Good pilot: draft a competitor report, build a briefing doc,
   triage leads into a dashboard.
   Bad pilot: anything that sends money, deletes records,
   or emails a customer without a human reading it first.

3. Which team runs it, and which product matches?
   Non-technical team, wants pre-built workflows out of the box
   -> Claude Cowork (more on-rails, vertical bundles ship configured).
   Team comfortable connecting plugins and building its own flow
   -> ChatGPT Work (more flexible, steeper setup).

4. Can you afford a week of unmeasured usage while you find out
   what the task actually costs?
   Neither product publishes per-task pricing. Week one is a
   measurement week, not a rollout week.

What the correct output looks like: a finished draft you’d have been comfortable handing a new hire on their first week. Plus a real number for how long the agent took against how long the task used to take you.

The failure mode: skipping the timing step, loving the output, and scheduling the workflow to run daily before you know what a single run costs against your plan’s usage. The exact risk both product pages warn about in the fine print: a team burning a month’s included usage in the first week of enthusiastic use.

Walk away with: a four-question filter for choosing your first agent pilot, and a rule for what a pilot is allowed to touch before you’ve measured it once.

The guardrail I’d never turn off

Both products ship with the same control: a plan you approve before work starts, and check-ins you configure along the way. Call it the Approval Leash: the agent can run for hours and touch a dozen tools, but nothing it produces goes anywhere real until a person signs off.

OpenAI’s own language undercuts a little of its confidence here. The company reports that in adversarial red teaming, its auto-review system “blocked 100% of attempts to extract protected data, including attacks the reviewing model had not seen during training.” That’s a genuinely strong red-team result. It isn’t a guarantee that a production system can never leak client data. Keep your own approval gate on anything that touches a client system anyway, regardless of what the safety layer claims.

The honest tradeoff on the Anthropic side runs the other direction: cost, not safety. Anthropic users have spent a chunk of 2026 complaining about high token burn on the newer frontier models. An agent that works for hours unattended is exactly the kind of workload that turns a complaint into a real invoice. Watch the usage meter the same way you’d watch a contractor’s timesheet.

Recap

Anthropic put Claude Cowork on mobile and web July 7. OpenAI answered July 9 by merging Codex into ChatGPT, launching ChatGPT Work, and starting to sunset the standalone Atlas browser. Both products take a goal instead of a prompt and hand back a finished artifact instead of a chat reply. Both bill by usage, not by seat, and neither has published what a task actually costs. “Work from anywhere” stopped being a slogan and started being a feature two different companies shipped in the same 48 hours. Pick one recurring task you already know the hours on, run it with approvals on, and keep the Approval Leash attached no matter which agent does the work.

FAQ

Do I need a developer on staff to use either of these? No. Cowork ships pre-configured vertical workflows aimed at non-technical operators. ChatGPT Work leans on its plugin directory and Plan mode, which takes a little more setup but doesn’t require code.

Which one is cheaper? Neither company has published per-task pricing. Both meter usage against your plan’s included allowance the way Codex already did, so the honest answer is: you find out in week one, not before.

Can I run this on client data? Both offer admin controls and audit trails built for that use case. Keep a human approval step on anything touching client systems regardless of the vendor’s own safety claims.

What if my team already uses Zapier or Make for automation? Nothing here replaces a working automation. These agents are best at the one-off or recurring analytical and drafting work a person currently owns, not at wiring two apps together.

Reply and tell me which one you’re piloting first, Cowork or ChatGPT Work. I’ll compare notes with you directly.

Andrew

P.S. If you want help wiring your first agent pilot into the tools you already run, the skill-building side of this conversation happens inside whop.com/abra-ai. Related reading: Google Rewired Its Productivity Suite on the same “AI moves into where work happens” pattern from a different vendor. More at muddventures.substack.com.