Who actually benefits when AI gets cheaper?

Cheaper intelligence is not the same as wider benefit

When the price of an AI model falls, three things can happen. A provider can lower its prices. A product can deliver more capability for the same subscription. Or the supplier can keep more of the difference. The outcome depends on competition, product design, and what customers are actually buying.

For someone trying to complete a useful task, model access is only one ingredient. They also need the right material, enough context to describe the job, a way to judge the output, and somewhere to use it. These complementary inputs can remain scarce while inference becomes inexpensive.

Our central question is therefore who gains usable capability when AI gets cheaper? We think the answer will depend increasingly on the services and institutions around the model: how work is organised, how knowledge is made available, and who can obtain reliable help when the output is wrong. That is a research proposition, not a claim that the price of computation has stopped mattering.

The price decline is real, but its unit is narrow

Stanford's 2025 AI Index reported a greater-than-280-fold reduction between November 2022 and October 2024 in the inference cost of a system reaching roughly GPT-3.5-level benchmark performance. This is a historical benchmark comparison, not a claim that every AI task or consumer subscription became 280 times cheaper. ^[1]

That distinction matters. A price per token measures a purchased input. A user wants a corrected spreadsheet, a finished design, a reliable translation, or a working piece of software. The input can improve dramatically while the cost of reaching the intended outcome changes much less.

Lower inference costs can still be transformative. They may make repeated checking, more generous usage allowances, or previously uneconomic tasks possible. But the path from a cheaper model to a better service has to be examined. It cannot be assumed from the model price alone.

Productivity depends on the work and the worker

Published studies show why a universal productivity claim is difficult. In Generative AI at Work, Brynjolfsson, Li, and Raymond studied a conversational assistant used by 5,172 customer-support agents. They found a 15% average increase in issues resolved per hour, with larger gains for less experienced and lower-skilled workers. The results came from an assistant embedded in an existing support workflow, with a defined task and organisational context. ^[2]

METR's early-2025 randomised study examined a very different setting: 16 experienced developers completing 246 tasks in repositories they knew well. Access to the AI tools available in that study increased completion time by 19%. This finding describes that population, tool period, and task mix. It is not evidence that AI generally makes developers slower. ^[3]

The update is equally important. In February 2026, METR explained that its follow-up experiment had become difficult to interpret because developers unwilling to work without AI were selecting out, payment changes could affect participation, and concurrent agents complicated time measurement. The team considered greater speedups plausible but warned that those data were weak evidence for their size. ^[4]

These studies should not be averaged into a single number. Together they motivate a better question: what has to be present around an AI system for a particular group to benefit? They also show why user enthusiasm, benchmark performance, and measured task completion belong in separate columns.

The access barrier can move

Imagine that a small organisation can now afford a capable model. It may still lack an organised knowledge base, permission to use the necessary data, a reviewer with time, or a workflow that accepts the result. The scarce input has moved from model access to the ability to use it well.

This creates two plausible effects. AI could reduce the advantage of prior expertise by making useful guidance available within the task, as the customer-support study suggests in its setting. It could also reward organisations that already have clean data, reliable processes, and people who can evaluate the output. Which effect dominates is an empirical question.

There is a distribution question inside the organisation as well. A worker may complete more tasks without receiving higher pay or more control over their time. A customer may receive a quicker response without a lower price. A provider may improve its margin. These are different beneficiaries of the same technical improvement.

For JDGMNT, making capability more accessible should involve recording who receives the benefit. Useful measures include the work someone can now complete, the effort they still bear, and whether the improvement persists after introductory assistance ends. Adoption counts alone cannot establish that.

FIG. 02 / THE HUMAN CONSTRAINT

The model got cheaper.
The hard parts didn’t vanish.

A simple model shows the limits of a price cut

Suppose an organisation's hypothetical first-month workflow costs are $120 for setup and learning time, $12 for the service, and $18 for human review. The total is $150. A 90% cut in the service charge saves $10.80, or 7.2% of the whole amount. The $120 is an assumed value for time, not a measured market price or necessarily a cash payment.

In the next month, if setup is not repeated and review stays at $18, the same service price cut reduces the assumed monthly total from $30 to $19.20: a 36% reduction. The same model-price change has a very different effect depending on the stage of adoption.

This example assumes unchanged task volume, quality, and review effort. A more capable or cheaper system could change all three. Setup may recur when tools change. Review may shrink with experience or rise when a user delegates harder work. The point is to expose the mechanism, not to estimate a universal percentage.

This also suggests a different intervention. For an occasional user, reducing setup effort may matter more than an additional discount. For a mature high-volume workflow, recurring inference costs may dominate. A subsidy or product feature should be aimed at the actual constraint.

Compare interventions, not just models

We would study a bounded, checkable task that participants already need to complete. A useful initial candidate is reconciling a small supplier catalogue into a consistent product sheet, using synthetic records during method development. It has clear errors to detect and does not require consequential decisions about individuals.

The proposed comparison would hold the underlying model and task conditions constant while varying support: model access alone; access with a prepared workflow; and access with a defined review-and-recovery process. A separate comparison could vary the service subsidy. This would help distinguish the effect of cheaper access from the effect of making that access usable.

Participants would be grouped or randomised under an appropriate study design, with prior experience and relevant device constraints recorded. The primary outcome would be a completed result judged against a predeclared standard by a reviewer unaware of the assigned condition where feasible. Record time, out-of-pocket spending, review effort, and failed attempts as separate outcomes.

People who decline, stop using the tool, or need help should remain visible in the study accounting. Reporting only successful users creates a particularly misleading picture when the question is access. Follow-up should also examine whether people can complete a similar task without the same level of assistance.

The result should change what we build

If a prepared workflow improves completion while a price subsidy has little effect, we should invest in task design and shared resources. If cost remains the binding constraint after the workflow is usable, pricing and delivery deserve attention. If improvements disappear once review effort is counted, the claimed gain needs to be reconsidered.

If experienced participants benefit and newcomers do not, a product positioned as widening access needs a different design or a narrower claim. If the gains are broad but accrue entirely to intermediaries, we should examine how pricing, ownership, and service commitments could share more of the value.

Our position is that falling inference costs create an opportunity to widen capability, but do not determine the outcome. The research challenge is to identify which complementary investments turn that opportunity into practical benefit, for whom, and at a cost that can be sustained. That is where JDGMNT can make a more useful contribution than repeating the next benchmark announcement.