OpenAI Sells an Agent at the Price of a Chat Window

Twenty dollars a month has bought a chat window since late 2022. On Nov. 20 OpenAI began using the same price to sell something that goes and does the work instead.

The company started pushing Dots, a personal agent, to ChatGPT Plus subscribers, the $20 tier. It is meant to take on a calendar, an inbox, and the handoffs between applications that a person would otherwise perform by hand, clicking and copying and checking.

OpenAI calls the release a test, a word it uses when it wants room to pull features back. Access is staged, and the company has not said how many subscribers are in the first wave or how often the agent may run unattended.

The installed base is what makes the tier worth defending. OpenAI said in October 2025 that about 800 million people use ChatGPT each week, most of them on a free plan that costs the company money, and the $20 tier is the first group that pays. Moving that group from questions to tasks is the cheapest growth available to a company with no advertising business.

The promise of a consumer agent has outrun the delivery for two years. Demonstrations are convincing because they show one task completed once. In production the error rate compounds: an agent that is right 95% of the time on each step is wrong more often than not across fourteen steps, and every one of those steps touches a real inbox.

The API told a different story the same day. GPT-Live-1.1 arrived with a billing change, cheaper standby pricing in place of long-lived connections, which matters most to developers running voice agents and live translation.

The old meter ran on time connected. A voice assistant kept a socket open for the length of a call and paid for the whole session, including the pauses while a person thought about what to say. Billing the standby state at a lower rate moves the cost toward actual use, and for a support bot that is the difference between a viable product and a demo.

Idle time is where real-time applications have been losing money. A model that answers in 300 milliseconds still spends most of a call waiting, and a customer who keeps a session open for eight hours while using it for ten minutes has been an expensive customer until now. Voice is the product OpenAI has most wanted to sell to call centers, and the price of silence was the reason several pilots stayed small.

Terms of service drew their own attention. Digiday reported that OpenAI revised the wording in its ChatGPT for business offering, after outside questions about how far its use of customer data reaches. Enterprise buyers read that language as a proxy for whether their documents end up in training runs, and they negotiate against the words rather than the product.

The business tier runs at $25 per seat on annual billing and $30 month to month, which puts the agent test in an odd position: the consumer product now promises more autonomy than the corporate one has been cleared to deliver. Procurement teams tend to notice that gap.

Competition has pushed every vendor toward the same feature. Google has been adding desktop control to Gemini, Anthropic has shipped computer use for developers, Microsoft has put agents inside Copilot and its own desktop, and Amazon has rebuilt Alexa around tasks rather than answers. None of them has published a price per completed task, because none of them can measure one yet.

Reliability is the constraint that none of them has solved. An agent that books a meeting needs calendar permissions. An agent that sends mail needs scope to send, and a wrong message cannot be recalled. The permission dialogs that follow are where adoption slows, and where enterprise deals stall in security review.

The economics are unfamiliar too. A chat subscription is priced against a few dozen messages a day. An agent that plans, retries and checks its own work can make a hundred model calls to finish one errand, and the margin on a $20 plan depends on how often subscribers use it. Flat rates assume light users who pay every month and forget; agents are used by the people who notice the bill.

Regulators are still writing the rules for software that acts. European transparency obligations and a patchwork of state laws cover automated decisions, but few of them anticipate an assistant that operates a personal inbox on behalf of someone who is not watching. An agent that sends mail under someone’s name adds a disclosure question too, since the recipient cannot tell whether the words came from the person or from a model trained on their writing.

OpenAI has not said when the test ends, what it costs to run per task, or whether the agent reaches the free tier. Those three numbers will decide whether $20 buys a colleague or a demo.

Related Posts

  • November 20, 2026
  • 15 views
Sandisk Completes Eight-for-One Split as NAND Outlook Darkens

Eight for one. Sandisk’s stock split took effect on Nov. 19, and the shares began trading on the adjusted basis on Nov. 20, a change that divides the price by…

  • November 20, 2026
  • 13 views
A Lawmaker’s Privacy Suit Puts Model Design on Trial

For three years British courts have sidestepped the question of whether the company that builds an AI model answers for what its users produce with it. A written defence filed…