Within a single week in July, the two companies that define the frontier of generative AI stopped pitching chatbots and started fighting over something bigger: your entire workday. Anthropic pushed Claude Cowork onto mobile and the web, and days later OpenAI unveiled ChatGPT Work, folding its Codex coding agent into a unified desktop app and rebranding the assistant as something closer to an autonomous colleague. Neither is a smarter answer box. Both are bets that the next interface for knowledge work is an agent that plans, acts across your apps, and hands back finished output.
For founders, marketers and operators, the shift matters more than another model-benchmark cycle. The question is no longer “which chatbot writes a better email,” but “which platform do I trust to run a multi-step project across my documents, inboxes and CRMs while I look away.” Here is what actually launched, how the two strategies diverge, and where the caveats bite.
What launched
On July 9, 2026, OpenAI launched ChatGPT Work, an agentic layer built to take action across a user’s apps and files, stay on a task for hours, and return completed artifacts such as spreadsheets, slide decks, documents and small web apps. It runs on the newly released GPT-5.6 model family, which OpenAI describes as tuned for reasoning through multi-step tasks, according to US News and The Next Web.
The structural move underneath the launch is the merger of OpenAI’s standalone Codex app into a single unified ChatGPT desktop application for Mac and Windows, with the previous app renamed “ChatGPT Classic,” as reported by Neowin. OpenAI framed the release as moving ChatGPT away from a chat interface and toward what several outlets characterised as a unified “super app” or full computing environment, per The Decoder. The company noted more than five million people use Codex weekly, with over a million already using it for tasks outside software development, according to The Next Web.
Anthropic reached the same destination from a different door. Claude Cowork — an agent for general knowledge work, not just coding, that shipped as a desktop app earlier in 2026 — expanded to mobile and web in the same week, with the rollout beginning around July 7 for Max subscribers, according to TechCrunch and 9to5Mac. The headline change is that Cowork can now run long tasks in the background even when a laptop is closed or offline: a user can start a job at a desk, get status updates on a phone, and collect the finished output later, as Engadget described it. To be precise about our own coverage area — zoho.social reports on Anthropic, not for it — this is an availability expansion of an existing product, not a brand-new launch, and the two releases landed within days of each other rather than simultaneously.
Two roads to the same prize
The two products are converging on the same job but backing different strengths.
OpenAI is leaning on breadth of integration. ChatGPT Work ships with a “Unified Plugins Directory” that launched with roughly a dozen or more integrations — Google Drive, SharePoint, Slack, Microsoft Teams, Gmail, Outlook, Salesforce, Adobe, Zoom, LinkedIn, GitHub, Canva and Dropbox among them — which users can invoke with “@” mentions or let the agent select on its own, per The Decoder. Billing is usage-metered like Codex rather than counted as standard chat requests, and the agent rolled out first to Pro, Enterprise and Edu users on web and mobile, with Plus and Business to follow, according to The Next Web. The pitch is a single console that reaches into every corner of the modern office stack.
Anthropic is leaning on delegation and governance. Cowork’s mobile-first design treats the phone as a remote control for agents that keep working while you don’t — a bet that the value of an agent is measured by how comfortably you can walk away from it. Anthropic paired the expansion with usage data drawn from roughly 1.2 million anonymised sessions, which suggested the top category of Cowork use was business-process operating at 33.4% — pulling scattered updates into reports, building onboarding checklists, reconciling spreadsheets — ahead of content creation at 16.4% and software development at just 8.7%, per TechCrunch. The expansion started with the Max plan (priced at $100/month) as a beta, with more plans to follow, according to PYMNTS.
Analysis: read together, the more interesting fact is who these products actually compete with. Neither is really fighting the other chatbot. They are competing with the IDE, the browser, the email client and the office suite — the software that knowledge workers open all day. That usage split, where mundane business operations dwarf coding, is the tell. The commercial prize is not the developer niche where these agents were born; it is the far larger population whose job is simply “work in software.”
The caveats
An agent that reads your files and acts across your apps is, by definition, a larger attack surface and a larger blast radius when it errs. Three caveats deserve weight.
Security claims need independent validation. OpenAI highlighted an “Auto-Review” layer that uses its most advanced models to check important actions before they execute, and said that during adversarial red-teaming it blocked 100% of attempts to extract protected data, including attacks the reviewing model had not seen in training, per Benzinga. That is a strong claim — and, as The Next Web noted, it “remains OpenAI’s own and has not been independently verified,” per The Next Web. A 100% block rate against a red team the vendor commissioned is a starting point, not a security guarantee; prompt injection through documents, calendar invites and third-party plugins is exactly the terrain where agentic tools have been most fragile.
Reliability and oversight of long-running agents. The selling point — an agent that works for hours, or overnight while your laptop is shut — is also the risk. The longer an agent runs unattended across live systems, the more compounding errors, silent failures and unintended writes matter. Anthropic’s own framing of Cowork leaving a “follow-up email drafted but unsent,” per TechCrunch, hints at the current compromise: keep a human on the last irreversible step. Treat that as the default, not an inconvenience.
Lock-in. Both companies want to own the whole workflow, which is the polite way of saying both want you to route your data, your integrations and your habits through one platform. The more your business processes live inside ChatGPT Work’s plugin directory or Cowork’s projects and artifacts, the higher the switching cost later. That is not a reason to abstain — it is a reason to keep your source data, and the connectors, portable.
The India read
For Indian enterprises and knowledge workers, the launches land on fertile ground and a familiar tension. The category that dominated Cowork usage — business-process operating, the reconciling and reporting and checklist work that fills finance, HR and admin roles — is precisely the layer where India’s vast services and back-office economy runs. Agents that credibly automate that work are not a niche developer tool here; they touch the core of how a great deal of Indian white-collar output gets produced.
Opinion: the opportunity for Indian founders and operators is real, but the governing constraint is data security and governance, not capability. An agent with standing access to customer records, financial spreadsheets and internal email is a compliance question before it is a productivity one — think DPDP-era obligations, client contracts and sectoral rules for BFSI and healthcare. Enterprises that pilot these tools should start on low-stakes internal workflows, insist on human sign-off for anything irreversible, and demand audit trails before letting an agent touch regulated data.
For India’s SaaS ecosystem, the strategic read cuts two ways. On one hand, a unified agent that reaches into fifteen apps is a threat to point solutions whose value was a single workflow. On the other, every agent needs somewhere to act — the integrations, the domain-specific tools, the systems of record — and Indian SaaS has spent a decade building exactly those. The companies that thrive will be the ones that make themselves the trusted hands an agent works through, rather than the interface it replaces. The chatbot era was about answers. This one is about who gets to do the work — and, just as much, who gets to watch over it.
