Unit Cost Dominance: where the tool becomes the substitute

Substitution begins at a cost threshold, not a capability threshold

Every general-purpose technology reaches a point where it stops assisting the worker and starts replacing the worker. That point is economic. It arrives when the technology, together with whatever human oversight it needs to be reliable, completes a task more cheaply than the person who used to do it, at the same quality or better. Below that line, the technology augments. Above it, the technology substitutes.

For general-purpose cognitive labour, the line has been crossed for a substantial and growing share of professional tasks. That is the first premise of the Discontinuity Thesis, and the rest of the sequence stands on it. The claim is narrow. AI does not need to do everything. It needs to do enough professional cognitive work, with a thin layer of human checking, at a low enough unit cost, that standalone human production stops being competitive across a large part of the knowledge economy. Once that holds, the remaining premises follow from how competitive markets behave.

This essay sets out the condition and the evidence that it has been met.

Dominance requires three conditions, and the third carries the argument

Take any cognitive task. The human cost is the wage needed to get a worker to complete it. The AI cost is inference plus the human time needed to check, correct and integrate the output into something usable. Unit cost dominance holds when the second figure is lower than the first at equal or better quality.

This is not automation in the usual sense. Nobody has to leave the loop. The human role only has to shrink from production to verification, and verification has to cost less than production did. A senior lawyer reviewing machine-drafted contracts costs less than a junior lawyer drafting them from a blank page. An analyst checking a generated pitch deck spends less time than one building it from raw data. An architect approving generated code spends less time than one writing it. The human is still present in each case. What has changed is the role, from producer to checker. The worker who used to produce the output has moved from necessary to surplus.

Three conditions have to hold together. The quality condition says the AI output matches or beats the human output for the task. The cost condition says raw AI production is far cheaper than human production. The workflow condition says that once oversight, verification and integration are added back in, the total is still cheaper. Critics often concede the first two and contest the third. The third is where the economics lives, so it is where this essay spends most of its time.

Frontier models now match expert deliverables on benchmarked professional work

The quality condition is anchored by GDPval, OpenAI’s benchmark of AI systems against industry professionals across 44 occupations. It contains 1,320 specialised tasks drawn from real professional work products, with a 220-task gold subset. Experienced practitioners grade the outputs blind, in pairwise comparison against human deliverables. The metric is whether the AI output is preferred to the human one or tied with it. By the standards of AI evaluation the design is strong. The tasks come from working practice rather than puzzles, and the judges are people who do the work.

The movement has been fast.

The point is not which model leads. The point is that expert parity on benchmarked digital deliverables stopped being a forecast and became a measurement, reached by several frontier systems from more than one provider within a few months. For the tasks in scope, the quality condition for substitution has been met. Whether a given task stays in augmentation or tips into substitution now turns on the cost of verification and integration, not on what the model can produce.

The newest releases extend the frontier, on vendor-reported figures

The releases of July 2026 did not add new GDPval figures to the public record, but they pushed further on the agentic and professional benchmarks that sit beside it. OpenAI reports that GPT-5.6 Sol set new highs on Agents’ Last Exam, which tests long-horizon professional workflows across 55 fields, and on BrowseComp. A week later Moonshot AI released Kimi K3, an open-weight model, and its model card places it level with or ahead of the closed frontier on several measures.

All figures are vendor-reported. The GPT-5.6 Agents’ Last Exam score comes from OpenAI’s launch post. The comparison rows come from Moonshot’s launch post and model card. “n/r” means not reported in those sources. Anthropic describes Claude Fable 5 as state of the art on nearly all the benchmarks it tested.

Read these with care. The BrowseComp row shows why. OpenAI and Moonshot report different scores for the same model on the same benchmark, because each vendor runs its own comparison. No single number in the table should carry weight. The direction should. Three providers, one of them releasing open weights, now cluster near the ceiling of benchmarks built to test multi-step professional work. That makes the quality condition harder to attribute to one lab’s evaluation choices.

GDPval’s scope bounds the claim without weakening it

GDPval is strong evidence inside a defined boundary. It covers knowledge work that can be done on a computer. It leaves out manual labour, physical tasks that need embodied judgement, work that leans heavily on tacit knowledge, deployments that need proprietary software access, tasks involving personally identifiable information, and roles where the human conversation is the product.

That boundary scopes the thesis. It does not weaken it. The thesis does not say all labour can be substituted now. It says the benchmarked domain overlaps heavily with the cognitive work that carried middle-class absorption in the postwar economy. Once that domain reaches parity, the wage-demand circuit loses the scarcity it was built on. The work outside the benchmark is real and some of it will hold out longer. Whether it can absorb the displaced population at the scale a wage-demand circuit needs is the question the later essays take up. GDPval does not show occupation-level automation. It shows that the quality condition has been met for a large class of self-contained digital deliverables. The economic condition is met wherever inference, integration and verification together still cost less than human-only production.

The raw cost gap runs to two orders of magnitude

The gap between inference and professional wages is very wide. In its GDPval discussion, OpenAI says frontier models produced comparable deliverables roughly a hundred times faster and a hundred times cheaper than expert professionals. It also warns that those figures reflect pure inference time and API billing, and exclude the oversight, iteration and integration that real workplaces need.

The July releases point the same way, again on vendor terms. OpenAI prices GPT-5.6 Sol at $5 per million input tokens and $30 per million output tokens. Moonshot prices Kimi K3 at $3 and $15. Moonshot also reports that K3 autonomously reproduced a set of astrophysics results in about two hours, cross-checking more than twenty papers and writing over 3,000 lines of Python, in work its team describes as one to two weeks for an experienced researcher. That is a vendor’s account of a showcase task, not a measured average. It still illustrates the size of the raw gap.

OpenAI’s caveat matters, and the strongest form of the argument sits inside it. The model-only benchmark understates deployed quality, because firms do not ship raw output. They ship AI plus a reviewer. The model-only cost figure overstates the deployed saving, because review costs money. Both corrections lead to the same place. AI plus a verifier undercuts standalone human production unless verification ends up recreating the whole original job.

Only verification that recreates the whole job restores the human producer

The arithmetic is simple enough to state in full. Set the cost of the old human-only task at 100. Set raw inference at 1. The deployed cost then depends on how much human checking the output needs, expressed as a share of the old production cost.

The figures are illustrative. The shape is the point. The human-only producer becomes competitive again only when checking, integration and failure handling absorb almost the entire original cost. A critic has to argue that the verifier is in effect doing the old job a second time. Where that is true, the task has not crossed the threshold and the thesis makes no claim about it. Where verification is materially lighter than production, the standalone human producer cannot compete on price. The burden falls on the critic to show that review rebuilds the job rather than compressing it.

This is the workflow condition in numbers. It is what makes the premise economic and not merely technical.

Augmentation assumes a higher tier of work that no longer exists

The usual answer to these numbers is augmentation. AI takes the routine cognitive work, people move up to judgement, creativity and strategy, productivity rises and employment holds. That story has dominated policy discussion for years.

It was right about earlier waves of automation. It is wrong about this one, and the reason is structural. Earlier waves automated bounded functions. The factory took muscle. The computer took arithmetic. The internet took distribution. Each time, human cognition stayed the bottleneck for whatever the new machine could not do, and displaced workers could move into work where cognition was still scarce. The wage premium on knowledge work survived because no prior technology touched the cognitive bottleneck itself.

AI automates enough digitally mediated, economically valuable cognitive work to remove general cognitive labour as a mass scarcity. There is no higher tier with mass absorption capacity waiting for the displaced. The technology doing the displacing is the same technology that would be deployed at any tier above. A junior analyst cannot escape into senior analysis at scale, because senior analysis is now also done by AI plus a reviewer. Reviewing roles exist, but one reviewer can oversee output that used to need a team of producers. The feature that let earlier waves preserve mass employment, a large body of cognitive work the new technology could not perform, is missing.

The position of any human refuge is also unstable. AI capability improves with each model generation, so the senior work it cannot quite manage this year is the senior work it manages next year. Augmentation is a real phase. It is a phase and not a destination, and it shortens with each release cycle.

Verification inverts the old asymmetry and cannot absorb the displaced

The gap between the cost of producing work and the cost of checking it is the asymmetry the wage circuit was anchored in.

For most cognitive work, producing has been dearer than checking. A competent contract, analysis or module of code took years of training and hours per item to produce. An experienced reviewer could check it in minutes. The labour market paid for production skill because that was where the scarcity sat. Review existed, but it was the cheaper side and earned a smaller share of the premium. The middle-class cognitive wage was, in large part, the price of scarce competent producers.

AI flips the asymmetry. Production falls towards zero marginal cost and checking becomes the binding constraint. The consequence for labour is direct. The checking side always needed fewer people, because it was the lighter task. One reviewer could cover many producers when production was hard. That ratio does not reverse when the asymmetry does. A review workforce sized to check producers is, by construction, too small to employ the producers it no longer needs.

The hopeful reading is that labour simply migrates from making to checking, and checking expands until it absorbs everyone. The arithmetic rules this out. If checking took as long as making, there would be no saving and no reason to deploy. The business case for AI depends on review being cheaper than production. In practice one reviewer supervises the output of several AI instances. The ratio varies by domain. The direction does not. Review absorbs fewer workers than production released.

Verification itself thins with each model generation

There is a second problem. Much review work is itself open to AI. A system that can draft a contract can also check it against a list of standard clauses. A system that can write code can also catch syntax errors and common antipatterns.

The symmetry is incomplete, and the exceptions are worth stating precisely. Evaluation is harder than generation in some domains, because spotting what is missing from an output is a different task from producing it. Missed issues in legal review, unhandled edge cases in code and unstated client preferences in a proposal are all places where human review keeps real value, because the failure is latent rather than visible on the page.

So the compression runs at different speeds. It runs fastest where faults show in the output itself: code that will not compile, a summary that drops named entities from its source, a contract whose clauses contradict each other. It runs slower where finding the fault needs outside context, institutional memory or judgement about what should have been there. Even where review keeps meaningful value, it is not a stable home for displaced workers at scale. It is a smaller layer working at a higher level of abstraction, and the next model generation usually thins it again. The reviewer of this year’s models is doing what the producer will do next year, and next year’s reviewers are fewer, working one level up, with the same dynamic repeating.

This is where P1 differs from an ordinary productivity story. The human role in the loop is real. It does not scale with the work, and it does not hold at the same intensity over time.

Residual human niches are real but too small to carry a circuit

The serious version of the objection is that some work resists substitution. Care work that depends on deep emotional attunement. Physical work in unstructured settings. Roles where human presence is the point. Creative work at the level where originality is the measure. Work that carries legal accountability.

These categories exist. They are smaller than the displaced population. Care work at the scale of mass absorption needs either much higher wages for care, which requires political choices that have not been made and show little sign of being made, or a change in how care is delivered and funded, which is a successor-system question rather than a continuity one. Embodied work is bounded by the pace of robotics, slower than software but not standing still. Trust-based roles rest on social conventions that are shifting as AI-mediated interaction becomes ordinary. High-end creative work is a small market. Liability-bearing roles are partly protected by law, but that protection narrows as AI judgement becomes admissible.

General cognitive labour is losing its role as the mass scarcity that supported middle-class absorption. Some human cognitive work survives. A wage-demand circuit needs mass absorption, not scattered residual scarcity, and the exposed population is too large for the niches to hold.

Deployment data shows the verifier architecture in production

The cost table is illustrative. Deployment records are evidence, and by July 2026 several of them place the verifier arithmetic in production economics rather than projection.

Anthropic’s case study with Novo Nordisk reports that clinical study reports, which once took up to fifteen weeks across forty to fifty professionals, can now be produced in minutes by a team of three using the NovoScribe platform built on Claude. AWS’s description of the same deployment gives the same figures. Resource needs for device verification protocols fell by 95 percent. Patient documentation that took months with outside agencies now generates in under a minute. Reviewers report that the output increasingly meets the bar that used to require human authorship, and the platform has drawn positive feedback from regulators in a heavily regulated industry.

OpenAI reports a similar pattern inside its own walls. At the GPT-5.5 launch in April 2026 it said more than 85 percent of staff used Codex weekly, across engineering, finance, communications, marketing, data science and product. The finance team used it to review 24,771 K-1 tax forms totalling 71,637 pages, two weeks faster than the previous year. A go-to-market employee automated a weekly business report and saved five to ten hours a week. The communications team built and validated a Slack agent that handles low-risk speaking requests automatically and sends higher-risk ones to a human. By the GPT-5.6 launch in July, OpenAI was reporting that AI writes the majority of research code inside the company. These are a vendor’s claims about itself and should be read that way.

The communications example shows the structure most clearly. Low-risk work is automated. Higher-risk work goes to a person. That is not stable mass complementarity. It is the verifier architecture running. The production layer compresses, the review layer stays, and the headcount the task needs falls by an order of magnitude. Human oversight is not the rebuttal to substitution. It is how substitution is built.

One qualification belongs here. Breadth of use is still modest. The Federal Reserve’s Real-Time Population Survey records generative AI assisting 1.7 percent of total US work hours, up from 1.4 percent in late 2024, with 37.4 percent of employed adults using it for work as of the third quarter of 2025. Those are self-reported figures. They show that the threshold has been crossed on benchmarks and in leading deployments well ahead of economy-wide diffusion. P1 is a claim about the cost condition. How that condition spreads through workflows is the subject of P2.

Competitive markets select against human-only production

Once dominance holds for a task, the competitive logic follows. A firm that keeps human-only production pays more for the same output. Rivals using AI plus review produce it for less. The market price moves towards the AI-plus-verifier cost. The human-only firm either matches that cost, which means deploying AI and cutting its producer workforce, or loses share until it leaves the market. In contestable markets there is no third route that keeps both human-only production and the firm’s position.

None of this needs enthusiasm for AI. It does not need executives who want to cut staff or boards that rank margin above employment. It needs only that markets stay competitive. A firm that holds human-only production above the AI-plus-verifier cost is selected against, whatever its intentions.

That is why augmentation does not last. Augmentation is the interval in which firms have deployed AI but not yet reduced headcount, because review protocols are still being learned, liability is still being settled and cuts are politically costly. It ends when competitors finish their reductions and the firm has to choose between matching them and losing share. The pressure is the same one behind every earlier round of cost reduction. What is unusual about AI is the breadth of cognitive work it reaches, not the economics of its adoption.

Task dominance becomes workflow recomposition once models operate the interfaces

GDPval measures tasks. The standard defence is that tasks are not jobs. That defence held while tasks stayed locked inside workflows only humans could operate. The worker was never only a producer of output. The worker was also the integration layer, carrying work between inbox, spreadsheet, CRM, browser and codebase. That fragmentation of software was a moat around white-collar labour, and it is closing. OpenAI’s launch figures already cited show the trend on OSWorld-Verified, the standard test of operating real desktop software: GPT-5.4 at 75.0 percent and GPT-5.5 at 78.7 percent, against a published human baseline of 72.4 percent. Once models can both produce the output and operate the interfaces it moves through, task-level dominance becomes workflow recomposition.

That propagation is the second premise, and the next essay.

The floor the rest of the thesis stands on

This essay sets out the empirical floor. AI plus verification performs a substantial and growing share of professional cognitive tasks at lower cost and equal quality compared with human-only production. On benchmarked digital work the frontier has reached expert parity, and each generation moves further into the remainder. Augmentation fails as a long-run account because there is no higher cognitive tier with mass absorption capacity. Review roles do not scale to take in the displaced. The work that resists substitution is real and too small to sustain a wage-demand circuit.

The later essays close the remaining exits in turn. Interface Collapse shows how task-level dominance spreads through workflows. The Multiplayer Prisoner’s Dilemma shows why no actor can restrain it alone. The Sorites Collapse Principle and Categorical Recursion close the regulatory route. The Successor System and Drag Is Not Rescue show that the structural alternatives preserve consumption, not the wage-demand circuit. This essay shows why those exits matter. The circuit cannot rest on the assumption that AI will stay uncompetitive with human cognitive labour. The evidence now runs against that assumption.

What would refute this

P1 would fail if the workflow condition failed in practice. The clearest sign would be deployed verification, rework and integration costs converging on the full cost of human production across a wide range of professional tasks, so that firms gained little from substitution. A sustained stall or reversal in frontier performance across independent professional-deliverable benchmarks would also count against it, though no single benchmark is decisive.

The labour-market signature the premise predicts appears first as non-absorption rather than as a jump in headline unemployment: weaker entry-level hiring in AI-exposed work, falling junior-to-senior ratios in cognitive sectors, and productivity gains that do not reach real median wages. A sustained recovery across several of those measures would count against the thesis.

On the junior signature, the evidence currently conflicts. The Stanford Digital Economy Lab’s payroll analysis, revised in August 2026 on a larger dataset, puts employment of workers aged 22 to 25 in AI-exposed occupations 19 percent below where it would be had it kept pace with less-exposed peers. The gap has widened since it was first reported in August 2025. In May 2026, researchers at the Federal Reserve Bank of New York reported the opposite pattern in job postings. They found no divergence between junior and senior postings within highly exposed occupations, found that the relative fall in vacancies for exposed occupations began before ChatGPT’s release, and put fewer than 10 percent of workers and vacancies in highly exposed occupations. The evidence register’s working reconciliation draws on PwC’s 2026 Global AI Jobs Barometer, which finds early-career postings flat in highly exposed sectors and junior roles in those sectors seven times more likely to demand traditionally senior skills. If entry-level titles survive while their content is seniorised, postings could show no split while payroll does. That is a hypothesis, not a finding. The conflict is unresolved, and the evidence register tracks it. If payroll data comes to match the postings data, the junior signature weakens, and the case for P1 shifts more heavily onto deployment economics.