On July 9, OpenAI took Codex, its coding agent, and folded it into the ChatGPT desktop app, alongside a new thing called ChatGPT Work. Twelve days later the company said the combined product had ten million users. Ten million, in twelve days, for a product that did not contain a new model.
Two days earlier, Anthropic had announced that Claude Cowork was coming to phones and the web, so that a task you start at your desk keeps running in the cloud while you check on it from the parking lot. Same week, OpenAI shipped GPT-Live-1, a voice model that listens and talks at the same time and decides, many times a second, whether to interrupt you. And yesterday AMD and Anthropic announced a two gigawatt chip deal, which is the kind of number that used to belong to power companies.
I keep a running list of AI news for this blog, and July’s list has a strange quality. Almost none of it is about a model being smarter. It is about where the model lives.
The thing being sold changed
I wrote in Why SpaceX Paid $60 Billion for a Text Editor that the value had moved from the model to the harness, the scaffolding of loops and memory and tools that turns a text generator into a worker. That deal, by the way, is still pending. What July made obvious is the next step in the same direction. The harness is not the product either. The product is the surface: the place the work already happens, with the model climbing into it.
Look at what OpenAI actually did. It did not release a new coding model on July 9. It released GPT-5.6, yes, but the ten million users did not show up for a benchmark score. They showed up because the thing that writes code and the thing they already talk to all day became one window. Codex on its own had about five million weekly users in June. Put it inside ChatGPT and it doubled in two weeks. The intelligence was the same. The door was different.
Cowork on a phone is the same move from the other side. The model in Cowork is whatever Claude you already pay for. What changed is that a long job no longer requires you to sit in front of a laptop watching it. You hand it off and walk away. That is not a capability. It is a place.
Why the door matters more than the brain
A raw model is a rented commodity, and every serious buyer knows it. Any of these tools can call Claude or GPT or Gemini, and the labs will sell that intelligence to anyone with a credit card, at prices that have barely moved in a year. I did a price check on that earlier this month. The flagship costs what it cost in January.
When the ingredient is a commodity, the value goes to whoever controls the shelf. That is the pattern in every industry that has ever had a commodity input. Flour is cheap. The bakery on the corner with the line out the door is not. The interesting question about a bakery was never the flour.
The labs understand this, which is why they are all building surfaces now instead of only building models. OpenAI is turning ChatGPT into a workspace. Anthropic is putting Cowork everywhere you might be. Google is folding Gemini into Search and Workspace and Android so thoroughly that there is no place left to type a question that is not also a place Gemini answers it. Each of them is trying to be the door you walk through, because the door is the part you cannot rent from somebody else.
What a voice model tells you
GPT-Live-1 is the purest version of this. It is not, as far as I can tell, dramatically smarter than the voice mode it replaced. What it does differently is listen while it speaks. It can be interrupted. It decides, over and over, whether to keep going or shut up. That is not intelligence in the sense the benchmarks measure. It is manners. And manners are what make a surface usable.
I have written before that the interface is collapsing into a prompt, and then into nothing at all. A model that knows when to stop talking is a big step toward nothing at all. The product is not the words it says. The product is that you can have it in your ear during a phone call, or in your car, without the experience being maddening. Every ounce of that value is in the surface.
Two gigawatts of something
The AMD deal is the odd one out on the list, and it is worth sitting with. Two gigawatts of GPUs, coming online from 2027, with AMD putting up to five billion dollars into Anthropic. That is not a surface. That is the opposite, the deepest possible layer, the electricity and silicon underneath everything.
What it tells you is that the labs are being squeezed from both ends at once. The model in the middle is a commodity whose price cannot go up. So they push down, into owning the compute so the commodity costs them less to make, and they push up, into owning the surface so they capture the value the commodity does not. The middle, the model itself, the thing everyone spent 2024 arguing about, is becoming the least interesting part of the stack. It is the flour.
What this means if you run a business
Buy the surface, rent the model, and never confuse the two.
Concretely: when someone pitches you an AI product, ask what surface it lives in. If the answer is a new tab you have to remember to open, be skeptical. Tabs die. The tools that survive are the ones that show up inside the inbox, the phone, the document, the calendar, the job you were already doing. A tool that makes you go to it is fighting the entire direction of the industry.
And when you build something yourself, an automation or a small internal agent, put your effort into the door, not the brain. Pick a mid-tier model, wire it so you can swap it in an afternoon, and spend the saved time making the thing show up exactly where the work is. In the inbox. In the ticket. On the phone. The intelligence is a line item on a utility bill. The place it appears is the whole product.
Ten million people did not sign up for a model this month. They signed up for a window.