ChatGPT Work and GPT-5.6: OpenAI's Answer to Claude Cowork — The 2026 Guide for UK Companies

On Thursday 9 July 2026, OpenAI launched ChatGPT Work — an agent built into ChatGPT that can take action across your apps and files, stay on a project for hours, and hand back finished work: documents, spreadsheets, slide decks, even working web apps. It is OpenAI's most direct answer yet to Anthropic's Claude Cowork, and it ships alongside a new model family, GPT-5.6. This guide explains what ChatGPT Work is, how it works, what marketing, product and sales teams can do with it, how it compares with Claude, Gemini and Microsoft Copilot — and, just as importantly for UK companies, where its limits are, when it is overkill, and what the security risks really look like.

By Toni Dos Santos, Co-Founder, Spicy Advisory

What Is ChatGPT Work?

ChatGPT Work, announced by OpenAI on 9 July 2026, is an agentic workspace inside ChatGPT. Instead of exchanging prompts and replies, you hand it a goal — “audit last quarter's paid campaigns and give me a board-ready deck” — and it plans the steps, works across your connected apps and files, browses the web where needed, and comes back with the finished deliverable. OpenAI says it can stay with a project for hours if the task demands it.

Under the bonnet, ChatGPT Work combines the agentic engine of Codex — which OpenAI had already been repositioning for business tasks, as we covered in our guide to OpenAI's “Intelligence at Work” push — with the new GPT-5.6 model family. If that architecture sounds familiar, it should: it mirrors the approach Anthropic took with Claude Cowork, launched in January 2026, which we've written about extensively — from the complete guide to Claude for companies to three real Claude Cowork marketing workflows. OpenAI has now validated the category: the agentic workspace is how the major labs believe knowledge work will be done.

How ChatGPT Work Works

You don't need to be technical to understand the moving parts. There are five:

Availability: in the desktop app, ChatGPT Work is available on every plan from day one. On web and mobile it reached Pro, Enterprise and Edu plans on 9 July, with Plus and Business following over the next few days. UK users are in the first rollout wave.

GPT-5.6: Sol, Terra and Luna

ChatGPT Work ships alongside GPT-5.6, which replaces the single-model approach with three durable tiers. For a non-technical team, the tier names are the only jargon worth learning:

Paid users can pick the tier and set an effort level per task — the same dial-up-the-thinking pattern Anthropic introduced with the effort selector in Claude Opus 4.8. Sam Altman claims GPT-5.6 is 54% more token-efficient on agentic work than its predecessor, which in practice means longer tasks complete within your plan's usage limits. For comparison with Anthropic's frontier tier, see our Claude Fable 5 business guide.

The Claude Cowork Parallel — and Why It Matters

ChatGPT Work is, unmistakably, OpenAI's Claude Cowork. The building blocks map almost one to one: standing access to your files and folders, long autonomous sessions, connectors into your business systems, and finished artefacts rather than chat replies. Anthropic shipped this model of working in January 2026; OpenAI has now matched it and put its enormous distribution behind it.

For buyers, that's good news twice over. First, competition: pricing and capability will move fast on both sides. Second, transferability: everything teams have learned about working agentically transfers. The workflows we documented in 3 Claude Cowork workflows for marketing, the delegation patterns in our guide for UK marketing consultancies, and the skills-based approach from building 50 reusable Claude skills apply to ChatGPT Work with minor translation. The skill you are really building is delegating and reviewing agentic work — and that's tool-agnostic.

Use Cases: Marketing, Product and Sales Teams

Marketing teams

Marketing is where agentic workspaces pay off first, because so much of the work is research-to-deliverable:

• Campaign reporting on autopilot: connect your analytics exports and ad platform reports, and delegate the monthly performance narrative — the agent assembles the numbers, flags anomalies and returns a formatted deck.
• Competitor and market monitoring: a recurring task that browses competitor sites, pricing pages and coverage, and files a weekly brief.
• Content repurposing at scale: one webinar or whitepaper becomes a landing page, an email sequence and a month of social posts — reviewed by a human before anything ships.
• Live dashboards with Sites: instead of another slide deck, ask for a shareable dashboard — the same pattern we showed with Claude's live Artifact dashboards.

If your marketing team is starting from scratch, our AI training for marketing teams covers exactly this delegation-and-review skillset.

Product teams

• Feedback synthesis: point the agent at support tickets, review exports and interview notes; get a themed synthesis with verbatim quotes and a prioritised opportunity list.
• PRD and spec drafting: the agent assembles context from your docs and drafts the first version, so product managers edit rather than start from blank pages.
• Clickable prototypes with Sites: turn a concept into an interactive mock-up your stakeholders can actually click through — before engineering spends a sprint on it.
• Release communications: one changelog becomes release notes, an internal enablement doc and a customer email, consistently.

Sales teams

• Account research briefs: before a first call, the agent compiles company news, financials, tech stack signals and likely pain points into a one-pager.
• Proposal and quote drafting: connected to your CRM and templates, the agent drafts proposals that a rep reviews and personalises — hours become minutes.
• Pipeline hygiene and reporting: delegate the weekly pipeline review prep: stalled deals, missing next steps, forecast summary.
• RFP first drafts: the agent mines your past responses and drafts answers, with humans owning accuracy and commitments.

The caveat that applies to all three teams: an agent's output is a draft until a human has verified it. The failure mode isn't that agents produce bad work — it's that they produce confident, polished work with an occasional wrong number in it. Build the review step into every workflow from day one.

ChatGPT vs Claude vs Gemini Workflow vs Microsoft Copilot

Here is how the four main options compare for non-technical teams, looking only at what you get in the web and desktop products (no developer tooling):

For a deeper head-to-head on the two agentic leaders, our Claude vs ChatGPT for business comparison is the place to start — and if you run several of these side by side, read our guide to monitoring AI spend across Claude, ChatGPT, Copilot and Gemini before the licence sprawl starts.

Limitations: When ChatGPT Work Is Overkill

An honest assessment, because the launch coverage won't give you one:

• If your usage is chat, you don't need an agent. Teams that use AI for drafting emails, summarising documents and answering questions will see little benefit from standing file access and hours-long sessions. A standard Plus or Business chat plan — or the free tier — already covers that.
• Supervision isn't optional. Delegating a three-hour task doesn't remove work; it converts it into specification and review. Teams without a review habit end up shipping unverified output — or re-doing everything.
• Early-days friction is real. Sites is a beta; the web and mobile rollout is phased; agent tasks are rate-limited; and desktop app access to local files will be switched off by many IT departments until governance questions are answered.
• Cost compounds quietly. Agentic tasks consume far more usage than chat. If you already pay for Copilot or Workspace AI, adding another agentic subscription per seat needs a business case, not enthusiasm.
• Change management is the actual bottleneck. Around 70% of AI pilots never scale — not because the tools fail, but because nobody redesigns the workflow around them. Our UK SME AI adoption roadmap explains the stages most teams get stuck at.

Security and Risk: What UK Companies Must Weigh

The capability and the risk are the same feature: standing access. An agent that can read your files, inbox and CRM, browse the web and produce outbound-ready documents has a fundamentally larger blast radius than a chatbot. UK companies should weigh four risks before rollout:

• Data breach exposure. Every plugin you connect widens what a single compromised account — or a single over-permissive agent session — can reach. Client data, payroll, unannounced plans: if the agent can read it, an attacker who hijacks the agent (or an employee who connects a personal account) can exfiltrate it. Connect the minimum, scope permissions per team, and audit connector inventories quarterly.
• Prompt injection via the built-in browser. When an agent browses the web, malicious pages can plant instructions that redirect it — the classic path to leaking connected data. Vendor safeguards are improving, but no lab claims this is solved. Treat web-browsing agents with access to sensitive connectors as a high-risk combination.
• Confident mistakes at scale. A wrong figure in a chat reply wastes minutes; a wrong figure in an agent-generated client proposal, price quote or board pack causes commercial damage. The polish of “finished work” makes errors harder to spot, not easier. Mandate named human sign-off on anything that leaves the building.
• Regulatory accountability doesn't transfer. Under UK GDPR you remain the controller of personal data an agent processes — the ICO will not accept “the agent did it”. Standing access to email and files almost certainly warrants a DPIA, and the new data regime raises the stakes further — see our guide to the Data (Use and Access) Act for UK businesses. Employees connecting agents to company systems without approval is the new shadow IT; give them a sanctioned path before they build an unsanctioned one.

None of this argues for staying out. It argues for adopting the way regulated UK firms are already adopting Claude and Copilot: scoped pilots, named owners, review gates, and an audit trail — the governance bar we described in our Mills Review analysis is becoming the norm well beyond financial services.

What UK Companies Should Do This Month

1. Diagnose before you buy. Map your team's actual workflows and AI maturity — the free self-audit takes 20 minutes and tells you whether an agentic workspace is the right next step or an expensive distraction.
2. Pick one workflow per team, not a big bang. One marketing reporting workflow, one sales research workflow, one product synthesis workflow. Measure hours saved and error rates for four weeks.
3. Write the governance one-pager first. Which connectors are allowed, who reviews output, what data classes are off-limits, who owns the DPIA. Two hours now saves an incident later.
4. Train the delegation skill, not the tool. Tools will keep converging — ChatGPT Work proves it. The durable capability is specifying, delegating and reviewing agentic work, which is what our UK AI training programmes build role by role.