OpenAI moved ChatGPT Work, the version of its product built for long-running tasks at companies, from the desktop to the web and mobile on October 15, and limited the rollout to paying tiers.
The order in which access was granted says who the company expects to buy. Pro subscribers came first, then a lower-priced Pro tier, then Enterprise and Education accounts. The free and lower-cost consumer plans were left out, and enterprise customers get a two-week preview period during which the feature is off by default.
The same day, the company began pushing a newer model, GPT-6.1 Sol, to Pro users inside Work and its coding product. The model was released at a developer event on September 29, where the company described its capability as approaching that of its most advanced system while charging about a fifth of the price for standard input and output tokens.
Agents that complete multi-step work are the product category every large vendor is racing to define. The difference between answering a question and finishing a task is the ability to use tools, to recover from failure and to keep working after the first attempt goes wrong. Those properties matter more to a business paying for output than to a consumer asking for a summary.
Pricing is where the competition is being decided. Charging a fifth of the top-tier rate for a model described as close to the top tier compresses the margin on the most expensive calls and pushes customers up the volume curve. A business that moves its document drafting and data preparation onto a cheaper model spends the same budget on more work.
The ability to work across connected applications and files has not changed with this release. The product produces documents, spreadsheets, presentations and websites from material the user connects. That capability is what makes an agent useful to an organization and what makes it a governance question, because an agent with access to a company’s files can be instructed to send data somewhere by a prompt the company never wrote.
Turning the feature off by default for enterprise accounts is a deliberate choice. Administrators who must opt in keep control of the rollout and can test a small group before enabling the whole organization. The default also shifts liability in a way that matters in commercial contracts, since the buyer decided to turn it on.
The tiers excluded from the launch are as telling as the ones included. Consumers on free and low-cost plans generate the largest volume of usage and the least revenue per user, and long agent runs cost more to serve than a quick answer. Keeping the feature behind a subscription is how the company avoids subsidizing heavy usage at the bottom of the ladder.
Distribution through mobile also changes how the product is used. A task that takes twenty minutes is not something a user supervises on a phone, which means the interface has to be designed around starting work, leaving and returning to a result. That is a different interaction than a chat window and a different measure of whether the product succeeded.
The competitive field is crowded with agents. Every large cloud provider sells one, a group of startups sells them as standalone tools, and enterprise software vendors are embedding them in the applications companies already run. What separates the offerings for a buyer is often the connections, the audit trail and the price per completed task.
Analysts said the constrained rollout is a capacity decision as much as a product decision. Long-running agent sessions consume computing capacity in bursts that are hard to forecast, and a company that opens the feature to every account risks the kind of outage that damages enterprise relationships. Releasing to paying tiers first is how a vendor learns the load profile before the demand arrives.
The cheaper model carries its own set of questions. A model described as close to the frontier at a fifth of the price implies either an efficiency gain in how the model is served or a capability gap that shows up on harder tasks. Buyers will find out which one it is by running their own evaluations, and the results will determine whether the cheap tier expands the market or moves existing spending down a price level.
Rolling a feature out by tier also gives the company a way to measure demand without committing capital. Usage data from the first group shows how long sessions run and which capabilities customers actually invoke, and capacity can be reserved against a distribution that has been observed rather than one that has been estimated from a waitlist.
The shift also changes who inside a company buys the product. A tool that drafts documents on a desktop was an individual purchase; an agent that runs for twenty minutes across connected files is a team purchase with an administrator, an invoice and a policy attached. Enterprise sales organizations are built for the second kind of buyer, and pricing tiers are how they are reached.
