- The rollout began on web and mobile for Pro, Enterprise, and Edu users, with Plus and Business tiers to follow over the coming days.
- OpenAI also launched a revamped ChatGPT desktop application that unifies Chat, ChatGPT Work, and Codex on both macOS and Windows, plus a hosted websites feature that lets users build and publish sites directly through the new agent — no separate hosting service required.
On Thursday, OpenAI drew a line in the sand. The company unveiled ChatGPT Work, a new AI agent that merges its widely used chatbot with Codex, its AI coding engine, to let white-collar workers build documents, presentations, and websites without writing a single line of code. Underpinning it all is GPT-5.6, the company’s most advanced model — and one that, only last month, the US government asked OpenAI to delay over national security fears.
The launch is many things at once: a technical milestone, a strategic gambit, and a clear signal that the race to dominate enterprise AI is accelerating faster than regulators can keep pace.
What ChatGPT Work actually does
The pitch is deceptively simple. ChatGPT Work takes the reasoning and code-generation capabilities that have made Codex a favourite among developers and wraps them in an interface that requires no programming knowledge. A marketing manager can ask it to build a client-ready presentation from a spreadsheet.
A product lead can have it spin up an interactive website to showcase a prototype. The agent handles the translation from natural language to working code entirely under the hood.
This is not OpenAI’s first agent product. The company previously offered Operator for task execution and deep research for investigative work, later consolidating them into ChatGPT Agent for individual users and Workspace Agents for enterprise workflow automation.
But ChatGPT Work represents something more ambitious: a general-purpose tool designed to sit at the center of a professional’s daily workflow, capable of reaching across connected apps, local files, and the web to produce finished deliverables.
The rollout began Thursday on web and mobile for Pro, Enterprise, and Edu users, with Plus and Business tiers to follow over the coming days. Alongside the agent, OpenAI launched a revamped ChatGPT desktop application that unifies Chat, ChatGPT Work, and Codex on both macOS and Windows, plus a hosted websites feature that lets users build and publish sites directly through the new agent — no separate hosting service required.
The model behind the agent
GPT-5.6 is the engine that makes all of this credible. OpenAI shipped it in three tiers: Sol, the flagship; Terra, a balanced mid-range option; and Luna, a lightweight variant optimized for speed and cost.
The pricing tells its own story. Sol costs $5 per million input tokens and $30 per million output tokens — aggressive numbers when stacked against Anthropic’s Claude Fable 5, which runs at $10 per million input and $50 per million output. Terra drops to $2.50 per million input and $15 per million output, while Luna bottoms out at $1 and $6. The three-tier approach is more than product segmentation; it is a deliberate attempt to make the cost argument as compelling as the capability argument.
“You can apply the model’s ability to code to solve problems across every industry,” said Ty Geri, product manager for ChatGPT Work, describing GPT-5.6 as “competitive with models that are far, far more expensive at twice the speed and much, much cheaper”.
Those capabilities extend well beyond coding. The model family demonstrates state-of-the-art performance across knowledge work, science, and cybersecurity — the last of which, it turns out, is precisely what made Washington nervous .
GPT-5.6 did not arrive on schedule. OpenAI had originally planned a broader rollout in June, but the Trump administration formally requested the company stagger its release, citing national security concerns around the model’s cybersecurity and autonomous agent capabilities.
Under an executive order issued in June 2026, the government asserted new authority to review frontier AI models before public release — a regulatory lever that had never been pulled so directly on a commercial product .
The delay was brief but symbolically significant. It marked the first time a major AI lab publicly acknowledged being asked to hold back a model at the government’s request, rather than voluntarily delaying. OpenAI received the green light to proceed in early July, and the full public launch followed on July 9.
The episode crystallises a tension that is becoming impossible to ignore: the most capable AI systems now sit at the intersection of commercial product, strategic asset, and national security concern.
Thursday’s announcement, with its emphasis on enterprise productivity and accessibility, can be read as OpenAI’s attempt to frame GPT-5.6 as a tool for office workers rather than a geopolitical flashpoint — even as the capabilities that triggered the delay remain very much intact.
The Anthropic Question
None of this happens in a vacuum. ChatGPT Work is a direct response to Claude Cowork, the general-purpose AI agent Anthropic launched on January 12, 2026, and has been iterating aggressively ever since.
Claude Cowork was the first major agent product to target non-coders specifically, combining Anthropic’s Claude models with the ability to work across a user’s local files, applications, and web browser to produce finished deliverables autonomously. In the months since, Anthropic has added plugins, persistent memory, and deeper enterprise integrations, steadily raising the bar for what an AI coworker can do.
The result is a two-horse race with enormous commercial stakes. Both companies are reportedly preparing for initial public offerings — Anthropic targeting as early as October 2026 with a potential trillion-dollar valuation, and OpenAI leaning toward 2027 after earlier signals pointed to a late-2026 listing.
Enterprise contracts are far more lucrative and sticky than consumer subscriptions, and the winner of the agent wars stands to capture a disproportionate share of that revenue.
OpenAI’s counterpunch is built on two pillars: price and distribution. The three-tier GPT-5.6 pricing undercuts Anthropic’s Claude Fable 5 by 50% or more at every level of capability, and ChatGPT’s existing user base — already numbering in the hundreds of millions — gives ChatGPT Work an install base that Claude Cowork cannot yet match .
Cost anxiety
Beneath the feature announcements and model benchmarks, Thursday’s event addressed a quieter but equally urgent concern: the sheer expense of using these tools at scale.
Enterprise adoption of AI agents has been tempered by sticker shock. Running autonomous agents that chain together dozens of model calls per task can generate bills that surprise even well-funded organizations.
OpenAI’s three-tier strategy is designed to let companies route simpler tasks through cheaper models while reserving Sol for genuinely hard problems — a pattern the company clearly hopes will make the economics of agentic AI feel manageable rather than terrifying.
Geri made the point explicitly, describing GPT-5.6 as competitive with far more expensive models. The subtext was clear: Anthropic may have had the agent market largely to itself for six months, but that market is about to get a price war.
