The Day the Marginal Cost of Knowledge Work Approaches Zero ── Extrapolating the pace of model generations from Opus 4.7 to Fable 5: how far could organizations, jobs, education, and geography change in 2027–2029?

Model generations used to turn over once a year; now it happens in months. What has changed is not only how smart the models are, but how usable they have become as agents (= a mode of use where the AI completes a multi-step job on its own, without a person directing every step). When the marginal cost of knowledge work (= the extra cost of doing one more unit of work) approaches zero, what changes first? This article walks through the scenarios in order of how likely they are to happen.

01Measuring the slope of generational change: from Opus 4.7 to Fable 5

Start with the timeline. The models from Anthropic (= the US AI company behind Claude) have moved through Opus 4.7, its refinement 4.8, the differently built Mythos, and now Fable 5. The gap between generations has shrunk from years to months. Look at what changed with each generation, and you notice the axis of progress flipped partway through. Early improvements were mostly about "smarts" — getting single question-and-answer exchanges right more often. But that is not where recent generations have gained.

What has grown is the ability to carry a long sequence of steps all the way to the end without human help along the way. The industry calls this agentic capability (= an AI moving through multiple stages on its own, from receiving an instruction to reporting completion). Raising single-answer accuracy by a few percentage points matters far less in practice than raising "the probability of finishing a 30-step job unaided." If a job has 30 steps, moving each step's success rate from 95% to 99% lifts the chance of finishing the whole thing from roughly 20% to roughly 70% (the difference between 0.95 to the 30th power and 0.99 to the 30th power). The value of the recent model generations comes down almost entirely to this one point.

Bring it into a pharma setting and the difference is easy to see. A few generations ago, the model could handle "summarize this paper." The current generation can handle "search the major papers from the past five years on this indication, narrow the candidates, cross-check the efficacy and safety statements, draft a promotional material with numbered citations, check it against our in-house style rules, and propose corrections." From one-off Q&A to taking on a whole piece of work. That is what the "slope" consists of. In literature work, by this author's estimate, a first-pass screening (= the initial sift that picks which papers are worth reading out of a large candidate pool) that once took half a day could fit inside an hour, with a person checking the results as it goes.

Let me draw a line here. Everything up to this point can be confirmed from public information; everything from here on is extrapolation (= extending a past trend into the future as a guess). There is no guarantee this slope holds through 2027–2029. Compute constraints, regulation, electricity costs, or a technical plateau could flatten it. But since several generations of improvement have stacked up in the same direction at the same pace, it is worth thinking through what happens if it continues. Read every section below as a conditional scenario.

02The economics of near-zero marginal cost: what gets cheap and what does not

Start with the term marginal cost (= the extra cost of one additional unit of work). The history of software is instructive. Building a program costs a fortune, but copying a finished one costs almost nothing. That is why music, maps, and dictionaries — anything copyable — saw prices collapse and industries reshape. What is happening now is the next stage. Not copying, but the additional unit of "fresh thinking" itself is approaching zero. One more research report, one more draft reply, one more language of translation. Through a pay-as-you-go API (= a gateway that lets software call the AI directly, billed by usage), that extra cost appears to be falling toward a few hundredths of a percent of the equivalent labor cost.

Not all knowledge work gets cheap at the same speed, though. What gets cheap is work whose output takes the form of text or data and whose correctness can be checked afterward. What does not get cheap is work where a person's presence is itself the value, and judgments where someone must accept responsibility for being wrong. Here is the asymmetry laid out.

In pharma terms: draft answers to medical inquiries (= questions sent in by hospitals and other healthcare institutions) get cheap; the approval stamp does not. You will be able to spin up dozens of draft materials, but the weight of the act of signing off — "this wording may go out into the world" — does not change. If anything, the more drafts pour out, the more review and approval become the bottleneck (= the choke point that sets the speed of the whole flow). The volume of work on the cheap side grows, and the load concentrates on the side that stays expensive. This pattern should show up most sharply in regulated industries.

One more point that tends to be overlooked: even if marginal cost approaches zero, fixed costs do not. Model fees, verification systems, safeguards against wrong output. It is the same structure as software: copies were free, but quality assurance kept costing money. So strictly speaking, knowledge work does not become free. The per-unit cost disappears and is replaced by the cost of maintaining the machinery that checks correctness. The changes to organizations, jobs, and education in the sections below all rest on this asymmetry and this substitution.

03The middle of the pyramid thins out: managers as information relays shrink

Break down a middle manager's job and the core of it is "translation": turning leadership's direction into the language of the front line, and turning front-line reality into reports for leadership — information relay up and down. Two other parts remain: "sharing responsibility" (explaining and owning decisions when something goes wrong) and "developing and evaluating people" (watching subordinates grow and deciding how to treat them). Of these, the first thing AI agents (= AI that carries out multiple tasks on its own from an instruction) will replace is almost certainly translation. Reading a policy document and breaking it into tasks for each person; rolling up weekly reports into an executive summary. Across the generations from Opus 4.7 to Fable 5, these tasks have already been reaching practical quality.

Once the translation function moves from people to agents, one assumption holding up organizational hierarchy collapses. The middle layer of the pyramid organization (= a hierarchy where headcount shrinks as you go up) originally came from a limit on information flow: one boss can directly oversee only a handful of people. If AI carries the flow, that limit loosens. There are already signs in the AI industry: reports of teams of a few people running development or content operations that used to take dozens grew more common through 2025–2026. Extrapolate the slope and by around 2027, the "small team, high output" organization is no longer an exotic case at a few frontier companies but an option considered as a matter of course. That is the first scenario.

This does not mean managers disappear. What changes is the content of the job. What remains is responsibility-sharing and people development. Sharing responsibility can only be done by a person, legally and psychologically. Pharma material review makes it plain: even if AI can flag every candidate deviation, the role of making the final call — "this wording may go out" — and answering for it when problems arise stays attached to a person's name. The same goes for development and evaluation: listening to a subordinate's doubts and talking through their career is not a relay task. A manager's day should shift toward less shuttling of reports and a higher share of judgment and conversation.

• Functions likely to shrink: translating information up and down, aggregating and reporting progress, coordinating meetings
• Functions likely to remain: accepting responsibility for decisions, developing and evaluating people, mediating between departments
• Functions likely to grow: supervising agents — checking their output and deciding how much to delegate

Here are the conditions under which this scenario fails. One is labor law. Under Japanese employment practice, redefining a job easily becomes a dispute over dismissal or reassignment, so organizational change lags the technical slope. The other is how fast companies internally accept the change. In an industry like pharma, built on documentation and audit readiness, internal rules recognizing "a report compiled by AI" as an official record have to come first. Between the moment something becomes technically possible and the moment the org chart actually changes, a lag of several years is the realistic assumption.

04Jobs do not vanish; the entrance changes: redesigning how experience is gained

How have junior staff learned the job? Take the minutes, write the first draft, clean up the data. Get it marked up by a boss, fix it, submit again. That loop was "the first rung of the ladder of practical experience." But that first-rung work is exactly what the current model generation does best. Minutes come from speech recognition plus summarization, first drafts from generation, data cleanup from agents. What follows is not the disappearance of jobs but the disappearance of the first rung. The footing on which juniors build experience goes away. That is the second scenario.

Look at where the first rung sits in each pharma role. For an MR (= medical representative; the role that provides healthcare institutions with information on the company's medicines), it was summarizing literature to prepare for visits and writing internal reports. For medical affairs (= the department that handles information from a medical and scientific standpoint), it was drafting abstracts and inquiry responses. For material review, it was the first-pass screening of promotional materials against checklists. All three look largely replaceable by 2026-era models. Juniors get to "stop doing" these tasks — but the instincts they used to acquire by doing them now have to be acquired some other way.

So companies will be pushed to redesign the apprenticeship on purpose. Training used to be embedded in the work; from now on, unless you build the training explicitly, it does not happen. Three plausible designs:

I will not make a call here on total employment; the range of forecasts is too wide. What deserves attention instead is the possibility of a reverse flow. The more AI output there is, the more demand there is for people who can verify it. In medicines, every generated sentence is tested against the Pharmaceutical and Medical Device Act (= Japan's law on medicines and medical devices, which includes advertising rules) and industry codes. Even as the cost of generating approaches zero, the cost of verifying does not — verification carries responsibility, and responsibility attaches to people. If the entrance for juniors moves from "making" to "checking," jobs do not vanish; the shape of the entrance changes. The way this scenario fails is if AI takes over verification itself and human verifiers become people who stamp approvals without reading. That fork depends on how far regulators and each company's audit function are willing to accept "verification by AI."

05Credentials and education get re-examined: from "knowing" to "being able to answer for it"

TOEIC scores, IT certifications, language and accounting qualifications. Most of these have functioned as proof that "I hold this knowledge." But looking at the model generations from Opus 4.7 to Fable 5, holding and reproducing knowledge is already becoming territory the model covers more cheaply and faster. Translation, writing code, reading financial statements. On the extrapolated price curve, an agent (= an AI that carries out work on a person's instruction) costing a few thousand yen a month takes over a good share of what these credentials certified a few years ago. On this extrapolation, the market value of "knowing" keeps falling.

One thing does not fall: proof of the ability to accept responsibility for a judgment. A pharmacist audits a dispensing, a physician decides a prescription, an audit firm signs financial statements. These are not proofs of knowledge; they are mechanisms that give legal backing to the declaration "if this is wrong, I answer for it." However accurate the AI's proposed prescription, the entity that bears the consequences is, by law, limited to a licensed human. The closer the marginal cost of knowledge work (= the extra cost of one more unit) gets to zero, the scarcer thing is not the work but the taker of responsibility — and the relative value of licenses, audits, and track records rises. That is the third scenario.

What matters here is the difference in how fast each type of credential changes. Licensed monopolies (= credentials without which the work is illegal: physicians, pharmacists, lawyers) cannot change roles without legislation. Diet deliberation and ministerial ordinances take years, so change is slow. By contrast, certificate-type credentials (= qualifications that confer no monopoly and serve only as a signal of ability) are priced entirely by the market, so their value erodes fast the moment AI starts offering the same ability cheaply. They share the word "credential," but they crumble at completely different speeds.

Education gets pulled the same way. If the value of drilling knowledge falls, the center of education should move to "training in verifying AI output and deciding what to accept." Pharma has a precedent: material reviewers are trained not to write copy but to spot deviations (= departures from the rules) in copy. This training of the "checking side" rather than the "making side" could become the prototype for education in the AI era. Can medical and pharmacy schools steer from national-exam cramming (= knowledge stuffing) toward verification training? That is the hinge of this scenario. If credential reform moves faster than expected, even the licensed monopolies' advantage wavers; if universities cannot reform themselves, a half-way future is possible in which the gap between education and practice widens while credentials survive as empty formalities.

06Cities, regions, and the ways this could fail: three conditions that break the scenario

One reason people gather in cities is that advanced knowledge services existed only there. Specialist lawyers, consultants, statisticians. This support was concentrated in Tokyo and Osaka, and small companies and hospitals in the regions were disadvantaged by exactly that physical distance. As the marginal cost of knowledge work approaches zero, part of this concentration dissolves. A 200-bed regional hospital gets literature-review support at the level of an urban university hospital for — on the extrapolated price curve — a few thousand yen a month. A regional manufacturer with 30 employees hands contract review of big-firm quality to an agent. This is not science fiction; it is simply the price and capability slope of a Fable 5-class model extended forward.

This does not mean cities hollow out. Something remains: trust and face-to-face contact. Accepting responsibility (previous section) ultimately comes down to relationships between people. Whom to ask for a signature, with whom to build a long-term deal — accumulated face time still drives those choices. The reason for urban concentration narrows from "the work is there" to "it is where trust is built." That is the fourth scenario. The commuting city shrinks; the meeting city remains. In pharma, the pattern resembles how the MR visit (= the representative who brings appropriate-use information on medicines to healthcare institutions) has shifted its weight from information delivery to trust building.

Finally, the conditions under which this whole article fails. All four scenarios above are extrapolations of one slope — "model capability keeps rising and prices keep falling" — and if the slope changes, so do the conclusions.

Forecasting is not calling a fixed future. It is extending the slope we can observe today (the pace of the Opus 4.7 → 4.8 → Mythos → Fable 5 generational turnover and the price trend) in a straight line, and writing down in advance the conditions under which the line breaks. The moment any of the three failure conditions becomes real, this article's scenarios are due for revision. What I ask of readers is not to swallow the scenarios whole, but to keep watching from your own post: is the slope holding, or has it broken?