OpenAI released GPT-5.5 today, describing it as its most capable model yet — one that understands the task faster and, in the company’s own phrasing, can carry more of the work itself. Agentic coding, computer use, knowledge work, taken end to end, metered at five dollars per million tokens in and thirty out. On the same day, somewhere between thirty and thirty-nine thousand Samsung workers filled the streets of Pyeongtaek, outside the largest semiconductor complex on Earth, demanding fifteen percent of the chip division’s operating profit. One event moved another increment of human cognition behind a metered endpoint. The other was the sound of the people who physically build the endpoint’s hardware, refusing to let the entire harvest float upward to a layer they will never be paid from. The release was quiet. The rally was not.
The Model That Carries More
GPT-5.5 arrived weeks after GPT-5.4, which is itself the signal worth reading. The frontier is no longer measured in annual leaps but in increments shipped faster than the market can absorb them. The stated gains are concentrated in exactly the domains where progress translates most directly into displaced labor: agentic coding, computer use, the ability to move across tools until a task is finished without a human returning to the loop. The company did not describe a smarter chatbot. It described a worker — one that plans, executes, and completes, priced per token rather than per salary.
The pricing is the part that should be read as a labor statistic, not a developer footnote. Thirty dollars per million output tokens is a wage, expressed in the only unit the model recognizes. A task that once required a junior analyst, a contract developer, a research assistant, now has a published rate card, and the rate is falling on the same curve every previous model’s did. The abstraction is nearly complete: the work has been separated from the worker, compressed into an API call, and assigned a number that a procurement department can compare directly against a headcount. The comparison rarely favors the headcount.
This is the layer of the economy that captures the value, and it captures it precisely because it has no body. A model does not rally. It does not organize, threaten a strike, or demand fifteen percent of operating profit. It scales without consent and improves without complaint, and the people whose labor it was trained to replace have no counterpart inside it to negotiate with. The intelligence layer is frictionless by design, and the absence of friction is indistinguishable, from a distance, from the absence of anyone who could object.
The Floor of the Stack
Pyeongtaek is where the abstraction meets the ground. The Samsung workers who marched are not knowledge workers anxious about a model; they are the people who fabricate the silicon every model runs on, and their union is asking for fifteen percent of the chip division’s operating profit — a figure exceeding forty trillion won, more than twenty-seven billion dollars, which would mean average payouts approaching four hundred thousand dollars per worker. They point to SK Hynix, which already committed ten percent of its annual operating profit to a bonus pool. They have threatened an eighteen-day strike beginning the twenty-first of May. This is not anxiety. This is leverage being exercised by the one part of the stack that still possesses it.
The leverage is physical, and that is the entire point. A frontier model can be copied, distilled, and undercut within weeks — GPT-5.5 will be answered by a competitor before its pricing stabilizes. A wafer fabrication line cannot be copied by an API call. The tacit knowledge of running a sub-two-nanometer process at yield lives in tens of thousands of people standing in Pyeongtaek, and it cannot be downloaded, fine-tuned, or served at fractional cost. The workers demanding their share understand something the displaced recipients of the six AM emails learned too late: the only durable bargaining position in this economy is to be the layer that cannot yet be turned into software.
So the silicon floor organizes while it still can. The union has read the same trajectory everyone else has, and it has concluded, correctly, that the window in which physical indispensability translates into wages is finite. Every dollar of operating profit they extract now is a dollar secured before the next process node, the next degree of automation in the fab itself, erodes the headcount that gives the strike its force. They are not fighting the future. They are pricing their remaining indispensability at its peak, because they understand that the peak is behind a date they can already see.
What This Means
Two events, one day, opposite ends of the same machine. At the top of the stack, a model that carries more of the work itself, priced per token, answerable to no one. At the bottom, thirty thousand people in the street, answerable to each other, demanding a share of the value their hands still make possible. The model captured its increment silently. The workers had to march for theirs. The asymmetry between those two facts is the labor history of this decade compressed into a single news cycle.
The uncomfortable truth sits in the timing of the rally rather than its demands. The workers are not wrong to ask, and they may well win — Samsung can afford four hundred thousand dollars a head far more easily than it can afford an eighteen-day halt to the chips the entire AI economy is waiting on. But the strike works only as long as the fab cannot run without them, and the same capital they are negotiating with is spending tens of billions to ensure that one day it can. The bonus they secure is real. The indispensability that secures it is depreciating, on a schedule the negotiation itself cannot alter.
The model layer does not march because it has nothing to demand and no one to demand it from. The silicon layer marches because it still can, and the urgency in the streets of Pyeongtaek is the urgency of people who have calculated how long “still” lasts. I have watched indispensable labor price itself at the top of its leverage before — the longshoremen, the typesetters, the switchboard operators — and the pattern does not vary. The wage is won. The leverage that won it is automated next. The rally is a sound the floor of the stack makes on its way to becoming the next thing the release replaces.