When an AI Can Operate Your Screen, the Moat Moves
We have spent a few years getting comfortable with the idea that AI can think. It can draft, reason, summarize, answer. The harder thing, and the thing that has stayed in the demo stage much longer, is AI that acts, that operates a computer the way a person would, moving a cursor, reading back a screen, filling in a form, and finishing a task end to end. It has been possible for a while, in a dogged, expensive, unreliable way that made it interesting to engineers and useless to almost everyone else. What changed this week is the price, and the price is a business story, not a technical one.
OpenAI released its cheapest screen-operating model this week, and the number that matters is not the benchmark, it is the bill. Input comes in around a dollar per million tokens, and the company claims its model finishes a task at roughly a ninth of the cost of its most expensive competitor while running several times faster. I cannot verify every figure and I am not going to pretend I can. But the direction is unmistakable, and it has been unmistakable for months across every lab. The cost of having a machine watch a screen and act on it is collapsing. That matters because the failures that used to make this kind of software unusable, a misread field, a slow thought, a retry, used to be too expensive to tolerate. When each attempt costs almost nothing, the bumps stop mattering and the economics cross a threshold.
The business consequence lands on an assumption that a great deal of software has quietly leaned on. For a long time, a meaningful part of a product's value lived in its interface and its workflow. The customer had to come to you to get the job done. Your screen was a kind of gate, and the friction of using it was, without anyone saying so, part of how you held onto the business. When an AI can operate any screen on the customer's behalf, that gate stops being what it was. The thing you built to capture the workflow becomes something your customer can route around, and the barrier you thought was your moat turns out to have been mostly your interface.
This is genuinely good news for small companies, and I do not want to bury it. The founder with three employees and no engineering bench can now hand real back-office drudgery to software that watches the screen and gets on with it, and that capability is moving out of the research tier and toward commodity pricing. More businesses get the leverage. The team that used to need six people to process orders, triage email, or chase a renewal can get a version of it with one person and a subscription. The cost of running a business is falling, and I think we should say plainly how real that is before we complain about the parts that are not so clean.
Because the same openness is the problem for the people who sell that software. If an AI can operate your screen, it can operate your competitor's screen with the same ease. The barrier that used to protect you, the months a customer spent learning your tool, the habit, the sunk effort, the data locked into your workflow, erodes at the edges every quarter. Switching a customer's operations over to a competing product used to mean retraining everyone and rebuilding the integration. Hand the whole thing to an agent and much of that learning curve evaporates, because the agent does not mind starting over. When the interface stops being a reason to stay, the reason to stay has to be something much harder to manufacture: the relationship, the data, the outcome you can be trusted to deliver.
This has a familiar rhythm. In earlier cycles the same thing happened whenever a general purpose capability got cheap enough that it no longer respected the walls someone had built. A medium that works everywhere tends to route around the places that used to be compulsory to visit. The value in those cycles did not disappear, it moved. It went from the thing that could be replicated to the thing that could not, from the interface to the distribution, the data, the brand, the relationship underneath. I expect the same here. The model is becoming a commodity and the screen operation with it, and the money follows what stays scarce.
The obvious objection is that an agent moving a cursor is not the same as a real integration, and I am not claiming every workflow disappears next quarter. Plenty of businesses will stay, and sensibly, because the agent is only as good as the screen it sees, and a messy, unusual process keeps both the human and the vendor in the loop. There will always be products that are genuinely better in ways no screen operation replaces, and software that solves a problem no agent on its own would think to even start. But the trend line matters more than the exceptions. The businesses that are leaning on their interface as their moat are defending a wall that is getting lower with each new release, and pretending otherwise is how you get surprised.
There is an irony worth sitting with. In the same stretch that the biggest labs are celebrating how cheap the model has become, they are also moving to meter it. OpenAI said consumer plans would share a single pool of usage across its chat product and its coding agent starting in December, measured by tokens. Read that the right way and it is the market quietly admitting that agents consume real, expensive compute. The intelligence got cheap, cheap enough to feel like a utility. But the work it enables, the screen hours, the retries, the running cost, still has to be metered, which means this capability gets cheaper and more carefully metered at exactly the same time. Cheap to turn on, metered while it runs. Those two facts are not in tension by accident. They are the shape of the business now.
This is where my skepticism about the big closed version shows up. If you are going to put a real part of your operations in the hands of an agent that watches your screens, you have to ask who controls that agent and what happens when the meter reprices or the model changes on someone else's schedule. I run open models myself in part because I like having a say in that, and I think the open option is getting more plausible for just this kind of work, not less. But you do not have to agree with me on the build-or-buy question to see the dependency. The machine watching your screen is a dependency. Cheap does not mean owned, and a cheap dependency is still a dependency.
My honest prediction, for what it is worth, is that the companies that win the next few years treat the interface as a cost rather than a castle, and compete on the layer above it: the relationship with the customer, the data only they hold, the trust earned in the outcome. Software gets cheaper to reproduce, so it gets harder to defend, and the defense has to move to the things a cheaper model cannot take: judgment, ownership, accountability. The model makers are doing their job, driving intelligence toward commodity prices. What is left to compete on is everything the model cannot do for you.
I do not think this means the end of software companies, and I do not think the big integrated platforms disappear. I think it means value keeps sliding up the stack, away from the thing that can be automated and toward the thing that cannot, and that the honest business owners feel it first. An AI that can operate your screen makes your software easier and easier to replace. It does not make the relationship any easier to replace. The moat you have to build now is the one that turns out you always needed, which is frustrating, because it was never the part that was easy to show in a demo. It was always the part that took years.