ChatGPT Work: OpenAI's enterprise play and the GPT-5.6 engine

01 What OpenAI announced: ChatGPT Work in one view

ChatGPT Work is OpenAI's new product for the workplace, introduced in the launch video embedded above on the company's own channel. The pitch is that instead of a general-purpose chatbot that employees happen to use at the office, ChatGPT Work ships with the things a company actually needs to run AI across a team: workspace connectors to internal tools, administrative controls, and agents that can carry out multi-step work rather than only answer questions.

Two names do most of the lifting in the announcement. Codex is the agentic engine, the system that plans and executes tasks like writing code, navigating files and producing finished artifacts. GPT-5.6 is the underlying model generation that powers reasoning and language. The pairing is the product: a model that understands, an agent that acts, and a workspace wrapper that makes both governable by an employer.

As with any launch, it helps to separate the demo from the deliverable. What is verified as of publication is the announcement itself, the product page and the launch video. What is not yet verifiable is how the agents behave across thousands of messy, real corporate environments. The gap between those two things is where this launch will be judged.

05 GPT-5.6: what the new model generation changes

Context for the version number: OpenAI's GPT-5 line, introduced in August 2025, collapsed the older split between fast chat models and deliberate reasoning models into a single system that decides how long to think about a problem. The numbered iterations since then have been refinements of that design rather than reinventions, and GPT-5.6 continues the pattern: better instruction-following, steadier tool use, and fewer abandoned multi-step tasks, which is precisely what an office product needs.

The practical changes are less about raw intelligence than reliability. Workplace agents fail not because the model cannot write a paragraph but because it misreads a spreadsheet schema, stops halfway through a task, or quietly does the wrong version of the right thing. The improvements that matter for ChatGPT Work are therefore measured in completed tasks per attempt rather than in leaderboard positions, a distinction OpenAI itself draws in its evaluation materials.

That framing is our interpretation of where the marginal gains now sit, and it is worth holding loosely. Public benchmarks still show meaningful movement on reasoning and coding suites, and a model generation can be both smarter and steadier. What no benchmark yet captures is whether the gains survive contact with a company's own documents and workflows.

09 Codex as the engine: from research demo to work product

The Codex name has a longer history than most people realize. It first appeared in 2021 as the model family behind GitHub's Copilot autocomplete, fine-tuned from the GPT-3 lineage on public code. OpenAI revived the name in 2025 for its agentic coding product, software that plans changes, edits files, runs commands and reports back. ChatGPT Work now generalizes that machinery beyond programming into general office work.

Calling Codex the engine is meant literally in the launch materials: it is the execution layer that turns a model's output into finished work. Where a chat model produces a draft, an agent produces a pull request, a formatted report or an updated ledger entry. The difference is the difference between advice and labor, and it is the entire economic premise of the product.

The open question is supervision. Agents that act need permissions, audit trails and failure handling, and the engineering around those is harder than any demo suggests. OpenAI's launch emphasizes administrative control as a first-class feature, which is the right tell. The company is selling trust as much as capability, and trust in software is earned in production, not in launch videos.

13 Why enterprises are the real AI battleground

Consumer AI monetizes attention and goodwill; enterprise AI monetizes budgets. A company that adopts a workplace product buys seats by the thousand, renews them annually, and pays at price points that consumer subscriptions rarely touch. That is why every major lab, OpenAI, Google, Anthropic, and Microsoft through its Copilot franchise, is converging on the same customer: the employee at a desk with a budget owner behind them.

Enterprise adoption has also moved from experiment to line item in roughly three years, as the chart below shows in approximate terms. The functions adopting fastest, engineering, marketing and support, share a trait: their output is digital and reviewable, which lowers the risk of letting an agent draft first. The figures are survey-based approximations and should be read as direction, not census.

There is a structural reason the workplace is the prize beyond budgets. Work tools are where proprietary data lives, and proprietary data is what makes a generic model genuinely useful to a specific company. Whoever holds the workspace integration layer holds the context, and holding context has been the moat of every enterprise software era since the database.

18 The competition: Gemini and Claude in the workplace

Google's counter-position is distribution. Gemini is woven through Workspace, the email, documents and meetings that hundreds of millions of people already use, which means Google can ship AI to offices without winning a single new login. Microsoft plays the same card with Copilot across its office suite, and it owns the default desktop on much of the corporate world. OpenAI's answer is to be the better product at the layer above the office suite.

Anthropic holds the other flank. Claude has strong traction among developers and technically sophisticated teams, Claude Code has become a serious tool inside engineering organizations, and Anthropic's enterprise positioning leans on safety and reliability. The competitive map in 2026 is therefore not one leader but three wedges: OpenAI's product velocity, Google's distribution, and Anthropic's developer loyalty, with Microsoft monetizing the incumbent suite regardless of which model wins.

Interpretation, offered plainly: the decisive factor is unlikely to be raw model quality, which is converging, but the cost of changing how a company works. Switching chatbots is trivial; switching the system that files your expenses, drafts your contracts and reads your backlog is a project. First-mover advantage in the workflow layer compounds.

22 Pricing, incentives, and switching costs

List prices across the category have settled into a recognizable band, roughly $25 to $30 per user per month for business tiers, with enterprise contracts negotiated upward for scale, compliance and dedicated capacity. Those are approximate public price points as of 2026 and they move often. The interesting economics are not in the sticker price but underneath it, where each seat must cover its inference costs, agent compute and the amortized cost of training the models in the first place.

The incentives to switch are therefore engineered, not incidental. Launches in this category reliably bundle migration tools, importers for existing documents, connectors to rival systems and admin dashboards designed to make the first ninety days feel free of friction. Every one of those features exists to attack the switching cost identified above, which tells you the industry itself believes lock-in, not technology, is the actual contest.

For buyers, the practical calculus is unglamorous: run a pilot on real work with a fixed evaluation set, measure completed tasks and error rates, and price the hidden cost of workflow change before signing. The launch video shows the ceiling of what the product can do. Procurement should test the floor.

27 What to watch next

The signals that matter over the next two quarters are measurable: seat adoption and renewal rates, the share of tasks agents complete without human rescue, and whether GPT-5.6's evaluation gains hold on private corporate data rather than public benchmarks. OpenAI has committed to publishing evaluation methodology for the new generation; the interesting comparison will be against identical tasks run by Gemini and Claude in the same workflows.

The distribution base is the reason this launch cannot be dismissed, whatever the product's rough edges. ChatGPT reached 100 million weekly users within two years of launch and, by OpenAI's own reported milestones, roughly 800 million by late 2025, the fastest consumer-software adoption curve on record, charted below. Converting even a fraction of that base into paid workplace seats is the entire thesis of ChatGPT Work.

Enterprise AI remains early enough that the honest conclusion is procedural rather than predictive. The launch is a claim about reliability that only sustained use can test, the competition is converging on the same customer from three different strongholds, and the buyer's best tool is still a boring pilot. The office, not the chat window, is where the next phase of AI gets decided.