The Week Everything Got Better and Nobody Felt It
Five developments arrived this week, each bearing numbers that admitted no argument, each adopted with the earnest enthusiasm of consumers who have been told — and have agreed — that faster, cheaper, and safer are the only adjectives that matter.
OpenAI has discovered what every luxury brand learned decades ago: the middle tier is where the money lives. On April 9, the company announced a one-hundred-dollar-per-month ChatGPT Pro subscription, slotting neatly between the twenty-dollar Plus plan and the two-hundred-dollar tier that preceded it. The new offering grants five times the Codex usage of Plus — ten times, temporarily, through May 31 — and access to GPT-5.3 Pro. Sam Altman marked the occasion by noting that Codex had reached three million weekly users, up fivefold in three months, and celebrated by resetting usage limits; he has promised to do this at every million-user milestone through ten million. The pricing ladder now extends from free to two hundred dollars in five increments, each removing a small friction, each widening the aperture through which the user may pour tokens and time. Anthropic offers a mirror structure. Neither company has published what, precisely, the limits are. The user is invited to pay more, and is assured it will be enough, and the assurance is renewed monthly.
The speed, at least, is no longer in question. Cerebras, the wafer-scale chip company whose hardware processes over a thousand tokens per second, demonstrated its Codex Spark platform building what it described as a working Salesforce-style CRM application in twenty-nine seconds. The demo, posted on X without ceremony, followed an earlier showing in which a functional browser game appeared in nine seconds flat. James Wang, writing on the Cerebras engineering blog, framed the competitive advantage as temporal: two identical companies, one shipping in six weeks, the other in three, are not in the same business. The CRM, of course, is not Salesforce; it is a scaffold, a proof that the scaffold can be erected before the architect finishes describing the building. Whether anyone will live in it is a question the demonstration was not designed to answer.
Google, meanwhile, has decided that understanding should look like something. The Gemini app now generates interactive 3D simulations embedded directly in conversation: molecular models that rotate, physics demonstrations whose variables the user adjusts in real time, charts that respond to the prompt "show me" as though showing were the same as explaining. Notebooks in Gemini package chats, files, and instructions into persistent projects synced with NotebookLM, creating what Google describes as a home for ongoing work. The language is domestic — a home, a place to return to, a relationship rather than a transaction. PaperOrchestra, from Google Research, deploys five agents to convert experimental logs into LaTeX papers with near-human citations. The question that sits politely in the corner of each announcement, unasked, is whether a simulation that can be manipulated is the same as a phenomenon that has been understood. It is, perhaps, the sort of distinction one stops noticing after the third or fourth rotation of the molecule.
Waymo has now accumulated 170.7 million rider-only miles across San Francisco, Phoenix, Los Angeles, and Austin, and the numbers are, by every conventional metric, extraordinary: ninety-two percent fewer pedestrian crashes resulting in injury, eighty-two percent fewer cyclist crashes, eighty-three percent fewer airbag deployments. The data has been peer-reviewed, published in Traffic Injury Prevention, and signed off by Mauricio Peña, Waymo's Chief Safety Officer. Five hundred thousand paid rides occur weekly. Two fatalities have been recorded in over 127 million miles, neither caused by the autonomous system. The cars do not drink, do not text, do not rage. They also do not drive on freeways — the human baseline against which they are compared does — and they operate exclusively in pre-mapped urban areas where every curb has been catalogued. The safety is real. The asterisk is small. The question is not whether the machine drives better than the human, but whether "better" is a category the bereaved will find consoling when the asterisk, one day, matters.
The most telling number of the week, however, required no engineering and no investment — merely a survey and fourteen hundred young people willing to answer it. Gallup's "Voices of Gen Z" study, Wave 3, surveyed 1,572 respondents aged fourteen to twenty-nine between February 24 and March 4. Usage is stable: fifty-one percent use AI weekly or daily, unchanged from last year. Excitement has collapsed: from thirty-six percent to twenty-two. Anger has risen: from twenty-two percent to thirty-one. Eighty percent believe AI will make learning harder. Only twenty-eight percent trust AI-assisted work; sixty-nine percent trust human-completed work. They use the product at the same rate and like it considerably less. This is not the profile of a technology being rejected. It is the profile of a technology being endured — adopted not because it delights but because the alternative, not adopting it, has been made, quietly, more expensive. The soma, as always, was never forced.