A person can now ask for a first explanation of an unfamiliar subject, a comparison of alternatives, or a draft of a working document without first finding somebody willing to spend an afternoon on it. That changes the price of getting started. It does not settle the price of finishing well.
That distinction is the starting point for this series. Which parts of information became cheap, and which costs remain? If we answer carelessly, we can mistake a falling software bill for a falling business cost. If we answer carefully, we can see where a small amount of money or attention might buy considerably more useful work.
“Almost free” describes a direction for some forms of retrieval, reproduction, and generated analysis. It does not describe verified facts, sound judgment, energy, permission to use somebody else’s material, or responsibility for an action. A useful explanation has to follow the whole chain from a question to an outcome.
Information has several prices
Consider a local repair business deciding whether to offer a new service. The owner needs to understand the equipment, estimate demand, compare suppliers, determine insurance requirements, choose a price, and communicate the offer. A search engine can locate relevant pages. A language model can turn them into a readable starting brief. Neither can make the customers appear or inspect the owner’s actual equipment.
The business faces a retrieval cost: finding potentially relevant material. It faces a comprehension cost: understanding what the material says. It faces a verification cost: deciding whether the claims apply here. It faces an execution cost: changing the business. Finally, it faces an error cost if the decision proves wrong.
Those prices can move in different directions. Retrieval may become cheaper while verification becomes more demanding because there are more plausible claims to sort through. A draft may become cheaper while the cost of correcting a misleading promise remains unchanged. More material can make a person better informed, or simply more occupied.
There is also a distinction between a copy and an original discovery. Copying an accurate explanation is different from paying for the experiment that made the explanation possible. Reproducing a supplier’s specification is different from independently measuring whether a particular shipment meets it. Cheap distribution does not erase the cost of establishing what is true.
This is why the price of a page, a token, or a subscription cannot stand in for the price of knowledge. Those are useful inputs to an estimate. The relevant unit is an accepted answer, a better decision, or a completed task with an acceptable error rate.
What the evidence actually supports
Stanford’s 2025 AI Index reported that benchmark-equivalent inference costs fell more than 280-fold between November 2022 and October 2024. That is a historical comparison for a specified capability level. It is evidence of a major decline in one computational price, not a measurement of every kind of intelligence or every business’s total operating cost.
The newer 2026 report continues to document capability changes. Neither report relieves a business of testing its own work. A benchmark result describes performance under the benchmark’s conditions. A business operates with old records, changing promises, incomplete inputs, and consequences the benchmark may not include.
The practical conclusion is narrower and stronger than “everything is free.” Some work that once required a substantial initial investment can now be attempted at much lower computational cost. Whether that attempt becomes useful depends on the surrounding process.
Suppose a model can summarize a maintenance manual quickly. The useful question is whether the summary preserves the warning that distinguishes two similar machines. A low inference bill would be irrelevant if a technician followed the wrong procedure. Conversely, a cheap first pass that helps a qualified technician find the correct page may be valuable without pretending to replace that technician.
That is the kind of boundary worth looking for: a cheaper step inside a process whose acceptance criteria remain explicit.
Cheap explanation changes who can begin
The first economic effect is access to a starting point. A person who lacks an analyst, a research assistant, or a large software budget may still be able to frame a question, collect a set of sources, and produce a rough comparison. The smaller the previous barrier, the less dramatic the improvement. The larger the previous barrier, the more meaningful it can be.
Imagine two businesses considering the same bookkeeping improvement. One already has an experienced operations manager who knows the relevant systems. The other has an owner trying to learn after closing the shop. A tool that produces a competent introductory explanation may save the first business little and give the second a useful way into the problem.
The improvement is still conditional. The owner must know what to ask, recognize gaps, and find authoritative material for consequential details. An explanation that is easy to obtain but impossible to assess can create false confidence. Access to a beginning matters; access to a trustworthy path beyond the beginning matters more.
This can change the range of projects somebody considers feasible. A small business might investigate a new category before paying for a consultant. A volunteer group might turn a confusing grant instruction into a checklist before asking a qualified adviser to inspect it. A student might compare several explanations until one becomes understandable.
In each case the gain comes from reducing a particular obstacle. It does not require claiming that the tool has become an expert on everything. It requires matching the assistance to the task and keeping a route to verification.
The cost shifts toward acceptance
When producing an answer is expensive, people ration answers. When producing answers becomes inexpensive, people need a better way to decide which answers deserve attention. Acceptance becomes a larger part of the work.
A purchasing comparison illustrates the change. Previously, the effort of assembling three supplier options might have consumed most of a morning. Now a tool can assemble a table quickly. But the owner still needs to check availability, shipping, warranty terms, compatibility, and whether the quoted prices are current. A clean table is a container for those facts, not evidence that the facts are correct.
An effective process gives every important field a source and a date. It marks missing values instead of inventing them. It separates a supplier’s statement from the business’s own observation. It records why an option was excluded. These practices make the comparison useful because another person can inspect the basis for the conclusion.
Without that structure, cheaper production can increase cleanup. Twenty attractive comparisons can be worse than one modest comparison that somebody has actually checked. The apparent abundance has been purchased with scarce review time.
This is related to the economic value of information discussed in The Gravity of Information. Information matters when it changes a decision in a way that improves the outcome. Its volume is a poor substitute for that effect.
A worked example: the price of an accepted supplier brief
Consider an explicitly hypothetical business that prepares ten supplier briefs per month. Before using AI, each brief takes ninety minutes to assemble and fifteen minutes to review. The total is 1,050 minutes, or 17.5 hours.
With an AI-assisted first pass, assembly takes twenty minutes and review takes thirty minutes. The ten briefs now consume 500 minutes, or about 8.3 hours. The direct reduction is about 9.2 hours. Suppose correcting errors and maintaining the instructions adds another two hours each month. The net reduction is about 7.2 hours.
These figures are an illustration, not an observed Salars result or a forecast for another business. Their purpose is to show what belongs in the calculation. The review burden doubled. Maintenance appeared as a new cost. The outcome still improved under these assumptions, but by less than the attractive assembly-time comparison suggests.
Now change one assumption. Suppose an unverified warranty detail creates a dispute that takes eight hours to resolve. That single incident can erase the month’s time saving. If it also damages a customer relationship, a time-only calculation understates the loss.
The appropriate response is not necessarily to abandon the tool. It may be to require direct verification of warranty language before a brief is accepted. The change targets the expensive failure while keeping the inexpensive assistance.
A second sensitivity matters: what happens to the saved time? If the owner has no useful work to move into that space, the business has gained capacity or rest rather than additional revenue. Both can matter. Neither should be reported as cash that was never received.
Some information remains expensive for good reasons
Certain facts require access, measurement, or professional responsibility. A property’s actual condition, a patient’s diagnosis, a product’s authenticity, and the contents of a signed agreement cannot be established merely by producing a convincing explanation. The relevant evidence may be unavailable to the model or unavailable altogether.
Private information also has a price beyond the computational bill. Customer records require appropriate handling. A business cannot treat confidentiality as an obstacle that cheap analysis conveniently removes. The ability to summarize a document does not establish permission to transmit it to a service.
Original research remains costly because it has to confront the world. Experiments require equipment, participants, time, and methods. An AI-generated hypothesis can help decide what to investigate. It cannot turn an unperformed experiment into a result.
Some information is deliberately scarce because the institution responsible for it must authenticate the record. A balance, license, court filing, or inventory count has value partly because a defined system stands behind it. Copying the text can be cheap while establishing its authority remains important.
Finally, knowledge can be local and tacit. A supplier’s history of keeping promises, the reason a customer rejects a particular design, or the sound that signals trouble in a machine may not appear in a searchable document. A system can help record and organize such observations. It should not assume them into existence.
Abundance can increase the price of attention
If everybody can produce proposals quickly, recipients receive more proposals. The sender’s cost falls while the recipient’s sorting cost rises. A tool can therefore improve one person’s productivity by imposing work on somebody else.
This is especially visible in outreach. A business may generate hundreds of messages for little money. The total value depends on whether the messages help the people receiving them, respect their preferences, and create conversations the business can actually serve. A higher sending rate is not automatically a better commercial system.
The same issue appears inside an organization. An employee who sends ten machine-generated summaries instead of one clear decision request may transfer the burden to a manager. The employee has produced more, but the team’s coordination problem has grown.
A useful rule is to charge generated output against the attention it consumes. Ask who must read it, what decision it supports, and what happens if it is ignored. If nobody can answer those questions, producing another document is unlikely to solve the underlying problem.
The next article, From Scarce Knowledge to Abundant Intelligence, examines this distribution problem more closely. Abundance at the point of generation is different from abundance at the point of useful application.
Lower prices can expand demand without improving every use
A cheaper resource often invites new uses. A business that once compared only large purchases may begin comparing smaller ones. A person who once asked one question may ask several. Some of these additional uses will be worthwhile; others will be marginal or distracting.
There is no contradiction in saying that a tool is valuable and that many uses of it are wasteful. The relevant question is whether the next use improves a decision enough to justify its total cost. A low per-use price makes that threshold easier to meet, but does not remove it.
This also explains why a falling subscription price need not reduce a business’s total AI spending. The business may process more work, demand higher quality, or introduce additional checks. It may choose to spend more because the capability is useful. That choice needs an outcome-based explanation rather than an assumption that cheaper units imply cheaper operations.
The strongest counterargument is that evaluating every small use can itself become burdensome. That is true. The response is proportional review: lightweight checks for reversible drafts, stronger checks for public promises and financial commitments. The process should spend attention where the consequences justify it.
The AI Leverage Equation develops this distinction into a practical estimate. Here the lesson is simply that unit cost, total cost, and net value answer different questions.
What could make this optimistic story fail?
Three failures deserve attention. First, a business may lack reliable underlying records. A capable tool cannot reconcile an inventory system that never records stock movements without somebody addressing the missing events. Better prose can hide a weak foundation.
Second, the accepted quality threshold may be higher than the tool can meet economically. A rough translation for personal understanding and a legally consequential translation have different requirements. Cheap generation may help with the former while adding an unnecessary intermediate step to the latter.
Third, the business may have a bottleneck elsewhere. Faster proposals are of limited value if production cannot fulfill accepted orders. Faster analysis can even worsen the queue by encouraging commitments beyond capacity. The correct intervention might be scheduling, staffing, or a narrower offer.
These counterexamples protect the useful claim from becoming a slogan. Falling information costs are an opportunity to redesign parts of a process. They are not a reason to ignore that process.
AI Leverage in Practice
Start with one recurring information task that has a recognizable finished result: a supplier brief, a meeting preparation note, a document comparison, or a customer-question draft. Write down what makes the result acceptable before changing how it is produced.
For a supplier brief, acceptance might require a source for every consequential term, explicit dates, no invented availability, and a separate list of unresolved questions. Measure the existing assembly and review time on a few ordinary cases. Keep those cases so the new process can be compared with the old one.
Use AI for a first pass on material you are permitted to provide. Require it to identify missing evidence. Have a person inspect the fields that could create a costly mistake. Record correction time as well as drafting time. If the process improves the accepted result without increasing consequential failures, retain it as a bounded capability.
Today’s tools can assist retrieval, summarization, comparison, and drafting. A future system might reliably acquire fresh authoritative evidence and complete more of the workflow. That possibility should remain a scenario until it is demonstrated under the relevant conditions. Do not purchase today’s reliability on credit from tomorrow’s promise.
Stop or narrow the trial if checking takes longer than the old process, if important errors survive review, or if the output has no clear recipient and decision. A useful stopping rule is part of responsible experimentation, not a sign of pessimism.
The useful meaning of almost free
Information becoming cheaper matters because it allows more people to begin, more alternatives to be considered, and more routine work to be attempted. Its practical value appears when the business can turn that cheaper input into an accepted result.
The remaining costs explain where the work goes next: reliable records, verification, judgment, permission, execution, and learning from consequences. Those are the foundations of The Age of AI Leverage, and the reason the series follows inexpensive cognition all the way into capital, business systems, ownership, and human purpose.
The question to keep is concrete: which step has become cheaper, who still has to accept it, and what must remain true for the outcome to improve?
Sources
- Stanford HAI, 2025 AI Index Report, historical inference-cost comparison, with its stated dates and capability threshold.
- Stanford HAI, 2026 AI Index Report, current report context; benchmark findings require task-specific interpretation.
Loading comments…