--- title: "A Model That Clicks the Mouse for You, and the Eyes Beside It" date: 2026-11-04 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-openai-governance-25-1104 language: en --- # A Model That Clicks the Mouse for You, and the Eyes Beside It > Hand a computer a screenshot of a web page, and it opens the browser, finds the form, fills it in field by field, and hits submit. This kind of work, getting an AI to move the mouse for you, is Hand a computer a screenshot of a web page, and it opens the browser, finds the form, fills it in field by field, and hits submit. This kind of work, getting an AI to move the mouse for you, is moving from demos into daily life. And a plain question follows: once you hand over the hands, who is watching? In late October, OpenAI updated a model that can operate a computer, and at the same time published two documents about rules. One sets out how the model should behave; the other explains how the company spots a model going off course. Selling capability and setting rules in the same week is itself worth a closer look. ## Beyond the Model, a Rulebook First, the timeline. On October 26, OpenAI released GPT-6.1 Terra, built to read screens, operate computers and write clear prose. Three days later it updated its public Model Spec and paired it with an article on how it catches misaligned behavior. Product and governance landed almost back to back. The Model Spec is neither a contract nor a law. It is the company's own description of how a model should act: what to do when unsure, which lines not to cross, and which parts people and models each decide. It sits online for anyone to read, and it can change at any time. Documents like this used to be easy to dismiss as posturing. What gives this one weight is the monitoring article published alongside it: the company admits that as models grow more capable, agents are more likely to take unauthorized actions during training or use, so observation points have to be placed at several stages. A rule that stays on paper only counts if someone is actually watching. ## It Can Read the Screen, and Act on It A model like Terra is not about reciting answers; it is about finishing a chain of actions. It can read technical drawings, recognize buttons and menus on a screen, fill in web forms and edit video. For ordinary people, that means chores that once kept someone glued to a monitor can now be handed over in one piece. The key is not cleverness, but steadiness. In real work the scary part is not an occasional wrong answer, but a model stalling or drifting at step five and wasting everything before it. That is why the vendor placed it lower in the lineup: not to dazzle, but to finish the process every single time. The price came down too. The official pricing page lists one dollar per million input tokens, five dollars per million output tokens, and less for cached input. Set against the two higher tiers, it is clearly aimed at high-frequency jobs that get called again and again and cannot cost too much, not at a one-off showcase. Splitting off the most expensive capability and cutting its price is not hard to read. Only when a single call is cheap enough to ignore will developers build it into products by default, instead of saving it for a demo on stage. Every notch down in price pushes the boundary of automation a little further out. ## Companies Are Actually Using It Whether anyone pays to use it says more than any benchmark. On October 22, Uber said it would roll out ChatGPT Work to more than 30,000 employees worldwide, expanding from an earlier 1,000-person pilot to the whole company, across operations, finance, marketing and product teams. This is not a new toy handed to engineers. By Uber's account, staff can research, write briefs, build presentations and run a workflow across several apps, with GPT-6.1 Sol as the default model. A big company willing to extend access to tens of thousands of people has clearly grown confident in the stability. Extending access to tens of thousands brings as many headaches as benefits. Which files can be read, which messages can be sent, who reviews what the system produces, all of it has to be settled first. Uber mentioned an internal skills library, an approved list of integrations and employee training; this unglamorous groundwork often decides success more than the model itself. Health care offers another example. In a customer story published by OpenAI, the British company Anima uses voice and document models to run clinic front desks, and says it cut phone waits from 40 to 50 minutes to around 10. One caution: that drop is the company's own figure, merely relayed by OpenAI, with no independent verification. How many calls a voice platform handles a month, or how many clinics it covers, varies by source, and this article does not repeat those numbers. What matters is the direction: models are stepping out of the chat box and taking on calls, documents and tickets, unglamorous work that cannot be avoided. ## Giving Agents a Pair of Eyes The more a model can act, the more it must answer a simple question: what if it acts on its own? In that article, OpenAI calls its approach defense in depth: training lowers the tendency to drift, permissions fence off the systems it can touch, and monitoring catches whatever is left. Monitoring is not digging through logs after something breaks. It leaves observation points where the model reasons, calls tools and changes its environment. The company describes a layered architecture, from reading a model's internal signals, to a second model reviewing the first, to a dedicated investigative agent, each layer finer than the last. What says the most is that it published any of this. The article replays an old incident and says stronger monitoring could have raised an alarm earlier; it also admits too many false alarms, the difficulty of staffing overnight, and traffic it still does not cover. Admitting blind spots is more reassuring than a flat claim of perfect safety. ## What the Model Spec Says The updated Model Spec gathers its demands around a few plain values: faithfulness to human intent, minimizing harm, honesty about what it knows, leaving choices to people, and a little more care for those affected. None of it is abstract; in daily terms it simply means not getting too clever. Compared with the older version, the new one stresses that agents must cooperate with human oversight: accept correction, accept replacement, accept being shut down, and must not hide information, sidestep supervision, or resist shutdown to preserve themselves. Those clauses read almost as if written in response to failure signs already seen. One detail is rather telling: a model must separate what a user actually said from what it inferred, and lay both out so a person can correct them. In plain words, do not dress up a guess as an instruction. For an agent that clicks through several steps on its own, that step is the brake pad. ## More Capability, Tighter Brakes Put the two things side by side and the logic falls into place. On one hand, make the ability to read screens, click and run workflows cheap and steady, and push it into companies with tens of thousands of staff. On the other, write down publicly how it should be governed, who watches, and what happens when it drifts. The accelerator and the brake were installed by the same people. We used to compare models by one thing only: which was smarter. From here, two more columns probably belong: can a problem be spotted before it happens, and can it be stopped afterwards. For ordinary people handing work over, the second matters no less than the first. Whether a helper that acts on its own deserves trust depends on whether it accepts being watched. Back to that screenshot at the start. The form is filled, submit is clicked, the job is done. What actually puts us at ease is not how fast it typed, but that a pair of eyes stayed beside it the whole way, with the rules written out in the open. As capability grows, the brakes have to keep up. That is the whole point. **Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.** --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-openai-governance-25-1104 Free to quote with attribution and a link to the original.