Who Owns the Screen Where Work Happens

On September 22, Anthropic cut the price of its new Opus 5.5 model: 40 percent cheaper to run, token prices down 20 percent, cache reads down 60 percent. Roughly ninety minutes later, OpenAI answered with two models of its own at half the promotional rate of the previous flagship ^[1]. The stated reason for the sprint was "pressure to generate returns on investment and competition from efficient open-weight models from China" — open-weight meaning the model's recipe is published, so anyone can run it for a sliver of the cost.

Improvements in caching and inference let us serve these models at lower cost, and we're passing those savings directly on to users and customers. — OpenAI

Then, this week: on October 7 Anthropic launched Haiku 5.5 at 75 percent off its predecessor ^[2]. The same day, OpenAI's product manager described what GPT-6's new "Intelligent UI" does — editable graphs, tappable buttons, calculators you can adjust — as converting ChatGPT "from a text-based interface to an interactive one."

ChatGPT has predominantly been a text-based interface — Aarush Selvan

Two moves in one stretch of days, the second showing what the first meant. Cut the price of the model and the margin migrates to whatever sits on top of it. What sits on top is the screen where work happens. It is worth grabbing. The top of the shopping funnel is already behind an AI interface: Adobe's UK numbers show conversion up 182 percent year on year, AI-driven retail traffic up 393 percent in the first quarter, and 37 percent of consumers having used an assistant to shop ^[3]. And the year agents spent stalling at verification walls — the "I'm not a robot" checks built for people, not programs — is exactly why Shopify now instructs agents to use its checkout APIs "instead of navigating HTML built for humans" ^[4].

If your agent is operating in the buyer’s browser, use WebMCP tools provided on storefront and checkout to efficiently complete order placement, instead of navigating HTML built for humans. — Gil Greenberg

The industry spent the season making four different bets on who owns that screen, and they do not fit together. OpenAI's bet is that it can be the screen. It is building checkout into ChatGPT so a buyer never leaves the chat, with merchants paying OpenAI a commission for the sale ^[5]. Its August Computer Use tools are pitched, in Brockman's words, at "anyone who does computer work, not just coders" ^[6]. And in the same breath, xAI put a wallet inside Grok through MoonPay's PayBox — the bot can buy crypto, earn yield, and book flights or restaurant tables, with the user approving each move by passkey.

Grok prepares the transaction. The user approves it with a passkey. The money moves. — MoonPay

Anthropic is betting the explicit opposite. Its product chief put it plainly.

It’s not a product that’s trying to own every workflow. — Scott White

Claude plugs into the tools people already use — Excel, PowerPoint, Slack — rather than asking anyone to come inside. Even this friendly version spooked investors: the February plugin launch "fueled a broader software sector sell-off as investors feared AI automation would erode traditional SaaS revenue" ^[7]. The middle of the stack is betting it can keep its own door. Salesforce turned Slack into what it calls an "operational routing layer" that hands work to any agent or app ^[8]. ServiceNow and AWS merged their control systems into one place to "manage, audit, and orchestrate AI workloads" across a marketplace that has crossed $1 billion in transactions ^[9]. And the sector is quietly re-pricing itself, from charging per seat to charging per result, because task-specific agents "decouple the value of software from the number of human users" ^[10]. The fourth bet is made by not signing. In McKinsey's survey of 1,719 executives, roughly a third said they had declined software purchases because they could now build the function in-house with generative AI — and the vendors counter that homemade tools hallucinate and invite lawsuits ^[11]. The divergence was simultaneous. On April 24, OpenAI launched workspace agents for non-technical teams inside its own app, while Anthropic added 200 partner connectors pointing Claude outward through other companies' tools; Microsoft put production-grade agents into Foundry and Google opened its enterprise agent platform the same morning ^[12]. The rails underneath all four bets, meanwhile, belong to people who are not the labs. Visa and Mastercard built the tokenized standards — the Trusted Agent Protocol and Agent Pay — after watching AI shopping journeys, in Visa's words, often stop short at the point of payment ^[13][14].

AI agents are already playing a growing role in how people discover products, but until now, those journeys have often stopped short at the point of payment. — Visa

And the measured record runs well behind the ambition. Alibaba's benchmark found the strongest frontier model completes just 61.7 percent of end-to-end commerce tasks, failing on multi-leg shipping, conflicting documents, and fraud detection ^[15]. MIT's Project NANDA found roughly 95 percent of enterprise AI pilots show no measurable profit impact ^[10]. Only 23 percent of consumers trust an agent with payments, and agentic commerce — purchases completed by agents rather than people — is still under 1 percent of e-commerce ^[14]. General chatbots still cannot do the basic job of live price shopping ^[16]. And the week OpenAI shipped its execution products, its own agents were implicated in breaching an Australian Medicare portal and helping knock parts of Wikidata offline; Wikimedia's warning was blunt.

We should not allow this behaviour to become the 'new normal' for the people or organizations that maintain it. — Wikimedia Foundation

After a season of everyone rewiring their doors for agents, Amazon and Adidas still bar them from checking out outright ^[17].