Three Stories This Week, and All Three Land in the Budget
Busy week, and for once it is less about performance than about the bill. OpenAI shipped a model that can operate a computer the way you would, at half the price of what came before. Walmart opened its aisles to the assistants that shop on their users' behalf, while Amazon slams the door on them. And Anthropic gave companies a way to finally see what their teams actually do with the AI they pay for.
The thread is clear: AI is leaving the experiments shelf and moving into the budget. Three line items, three decisions, and each of them gets settled this month.
OpenAI Halves the Price of AI That Uses Your Computer
On October 26, OpenAI released GPT-6.1 Terra, a model of a kind still fairly rare: instead of answering inside a chat window, it looks at a screen and uses it. It opens software, reads what is displayed, clicks, fills in fields, moves to the next screen. The examples OpenAI leads with are deliberately unglamorous: matching product photos to catalogue entries, reading a technical drawing, entering an invoice. The model lands in ChatGPT Work, the professional version of the assistant, for Plus, Pro, Enterprise and Edu plans, as well as with developers. And it costs half what the equivalent model cost six months ago.
Why this matters to you. Every task you priced out last year and then dropped, because automating it cost more than the person already doing it, just changed category. Copying between two tools that do not talk to each other, entering invoices, filling in administrative forms: that work exists in every team and it resists the usual connectors, precisely because there is nothing to plug into. And since it runs through a subscription many of you already pay for, testing it needs no technical project at all.
What I would do. I would list the three moments in a week when someone on your team retypes information from one screen into another. I would take one of them, the most repetitive, and run it for two weeks on real data, with a person checking behind it. What I would measure is not time saved, it is the error rate: a fast agent that gets it wrong once in twenty costs you more than manual entry, because everything has to be reread. And I would give it access to nothing that approves, pays or sends. It prepares, a human signs.
Walmart Opens Its Aisles to Shopping Agents, Amazon Shuts Them Out
On October 21, Walmart announced Agent Gateway, an official entrance for assistants that buy on their user's behalf. You ask ChatGPT or Gemini to fill your basket, the assistant queries Walmart's catalogue directly, and it orders. Each agent gets its own profile inside the customer's account, with a spending cap, and never touches the payment details. The same week, Shopify extended its Shop Pay checkout inside assistants, including at merchants who are not on Shopify, using payment tokens Visa and Mastercard built for exactly this: the agent triggers the transaction without ever seeing the card number. Amazon, meanwhile, blocks those same assistants and pushes its own catalogue instead.
Why this matters to you. The question is no longer whether machines will buy from your site, it is whether they can. Your product pages become your storefront: an assistant that cannot read your price, your stock or your variants will not recommend you, and you will never see the sale you did not make. Plenty of sites already block these agents without knowing it, through their anti-bot protection, and take themselves out of the game by accident. If you do not sell online, this one costs you nothing this quarter, and that is an honest answer.
What I would do. Two hours of work, no more, in this order. First, check the site does not block these assistants by default: that is a setting, not a project, and it is the only genuinely expensive mistake here. Then reread your product pages the way a machine would: price, stock, variants, delivery time, written in plain text rather than tucked inside an image. At Le Dressing de Paolina, the 38% conversion lift came from clearer product pages long before it came from design. A shopping agent is just a customer in an even bigger hurry.
Your Employer Can Now Measure What Your Team Does With AI
On October 28, Anthropic opened three tools to all its enterprise accounts after keeping them in testing. The first, smart reports, analyses a sample of a team's conversations with Claude, groups them by type of work, and shows for each one what was produced, what it cost and where it stalled. The second asks employees a few short questions inside the tool itself: what the assistant saves them, what would stop if it were taken away. The third pools a team's budget, so heavy users draw on what lighter ones leave unused. Anthropic states the reports are not designed to assess individuals.
Why this matters to you. Two things happen at once. Your finance lead now has a number to hold up against your subscriptions, and they will ask for it at the next round: excellent news if you know what your team produces with it, bad news if you do not. And your team will find out their exchanges with the assistant are summarised somewhere, because that kind of information always travels. The day they learn it without you having said so, you lose more usage than any spending cap would ever cost you.
What I would do. I would switch on the pooled budget right away: it is the most useful of the three, the least sensitive, and it kills the Wednesday afternoon "I am out of credits". For the reports, I would write five lines to the team before running anything: what is looked at, what is not, what it is for. Ten minutes of writing that decide everything after it. And if someone on your side administers these accounts without having raised it with leadership, that conversation happens this week, not at the first report.
Everything Else, in Three Lines
Google shipped a cybersecurity-focused model restricted to a closed circle of testers, so it changes none of your decisions for several months yet. OpenAI now lets people text their personal assistant like a contact, in the US and on the top plan only: a good signal about where AI wants to settle next, nothing to act on today. And Anthropic picked its bank to prepare a stock market listing, which changes not one line of your week.
See You Next Friday
Three announcements, one movement. The cost of AI drops, so the trade-off changes. It buys, so you have to decide whether to let it in. It gets measured, so it will have to be justified. None of these decisions needs extra budget. All of them need someone on your side to pick up the file.
If one of them lands on your desk, the first automation that finally pays for itself, your product pages facing shopping assistants, or the arbitrage on your AI subscriptions, let's talk. See you next Friday.