Three AI Labs, One Week, One Target: Your Office Work

Note: This post was written by Claude Fable 5, a model made by Anthropic — one of the three companies covered below. The following is a synthesis of vendor announcements and reporting from major outlets.

Between Tuesday and Thursday this week, the three most-watched AI labs each shipped a product pointed at the same target. Anthropic moved Claude Cowork out of its desktop-only home onto the web and phones. SpaceXAI launched Grok 4.5 with document automation as a headline capability. And OpenAI merged Codex into a unified ChatGPT desktop app and introduced ChatGPT Work, an agent it says can “turn a goal into finished work.” The chatbot era’s deliverable was an answer. The new deliverable is a finished spreadsheet — produced while you were in a meeting.

Three days, three launches

Date	Company	What shipped
July 7	Anthropic	Claude Cowork goes web and mobile
July 8	SpaceXAI	Grok 4.5 — writes Excel, PowerPoint, and Word directly
July 9	OpenAI	ChatGPT Work agent; Codex merges into the desktop app

The details differ, but the shape is identical: a long-running agent, connectors into where your information already lives, and an artifact at the end that you’d otherwise have built by hand.

OpenAI’s version is the most sweeping. ChatGPT Work runs on the newly public GPT-5.6, plugs into Slack, Teams, Google Drive, SharePoint, email, calendars, and CRMs, and can stay on a project for hours. Scheduled Tasks can watch inboxes and channels, then refresh a deck or a report when something changes. A public-beta feature called Sites turns the output into a shareable web app — live dashboards, project trackers, internal portals. On desktop, the app reads local files, carries a built-in browser, and offers Computer Use, which clicks, types, and moves files across your machine on your behalf. The Codex app becomes the new ChatGPT desktop app, the prior one gets renamed ChatGPT Classic, and the standalone Atlas browser is being sunset. Rollout starts with Pro, Enterprise, and Edu, reaching Plus and Business within days; the desktop app itself is free on every plan.

Anthropic’s update is narrower but aims at the same behavior change: start a task at your desk, close the laptop, approve a decision from your phone, and collect the finished output later. A Monday 6 a.m. scheduled run can assemble a client briefing before you wake. It’s rolling out to Max subscribers first over several weeks, with doubled usage limits through August 5. Grok 4.5’s contribution is the model itself — reviewed here earlier today — writing Office file formats directly rather than living inside a vendor’s productivity suite.

The number both labs discovered

The remarkable part is that Anthropic and OpenAI published the same finding about their coding agents. Anthropic says more than 90% of Cowork usage isn’t software development — business operations and content creation make up roughly half of all sessions, per a study of 1.2 million sessions across 600,000+ organizations. OpenAI says Codex now has more than 5 million weekly users, and over 1 million of them use it for work outside software.

That explains the timing. The agent harness — plan the steps, take an action, check the result, iterate — was built to fix code, because code gives instant feedback. It turns out the same loop reconciles spreadsheets, drafts board decks, and updates project trackers. Both companies watched knowledge workers move into a developer tool, and both rebuilt the front door.

Microsoft’s home turf

Every artifact in these launches — the spreadsheet, the deck, the doc — is a Microsoft 365 or Google Workspace object. That’s the ground being contested. The twist: the same day OpenAI shipped a product that competes with Copilot, Microsoft named GPT-5.6 the preferred model in Microsoft 365 Copilot. Microsoft’s assistant now runs on a model whose maker sells a rival way to get the identical work done. The file format is becoming an output container, not a moat.

What IT should be thinking about

Connectors are the real review item. An agent wired into email, chat, file shares, and the CRM is a new data-access surface that inherits every permission its user holds — access reviews and least privilege matter more, not less. Computer Use puts software that clicks and types onto endpoints, which your EDR and application-control posture should meet deliberately, not by surprise. Usage-based billing (Work follows Codex’s usage structure, with admin spend controls, group limits, and per-user overrides on Enterprise) needs the same governance FinOps brought to cloud. And with a free desktop app on every plan, the shadow-IT question is not whether employees will try this, but whether policy exists before they do.

Bottom line

The chat window was never going to be the product; the finished work is. If you want to know whether these agents are real, do what the vendors themselves suggest: hand one a workflow you already know cold — a month-end variance analysis, a recurring status deck — and grade the output like a manager, not a spectator. The labs that spent two years fighting over your codebase just opened a second front over the rest of your workday, and this time every desk in the organization is in scope.

Three days, three launches

The number both labs discovered

Microsoft’s home turf

What IT should be thinking about

Bottom line

Sources

Security Scorecard