For most of economic history, sophisticated intelligence has been expensive.
If a business needed difficult analysis, specialist knowledge, strategic thinking, software engineering or complex problem-solving, it needed people capable of doing that work. Those people were finite. Their time was finite. Expertise took years to develop and was expensive to employ.
That scarcity shaped organisations.
It influenced which problems were worth analysing, which decisions received senior attention, which ideas were explored and which opportunities were ignored. There were always more potentially useful questions than there were capable people available to answer them.
Artificial intelligence is beginning to change that constraint.
This was implicit in The Moment AI Changed: intelligence was starting to become something businesses could build with rather than something available only through human labour. What has happened since is not simply that models have become more capable. The economics of accessing useful machine intelligence have also moved extraordinarily quickly.
That combination matters.
As intelligence becomes cheaper and more abundant, intelligence itself may become a weaker source of competitive advantage.
The scarce resource moves somewhere else.
AI doesn’t eliminate scarcity. It moves it.
The cost curve is moving remarkably quickly#
The rate of change is easy to underestimate because most attention goes to model capability.
But capability is only half of the economic story. The other half is what that capability costs to access.
Stanford’s 2025 AI Index gives a striking illustration. The estimated inference cost of a model achieving roughly GPT-3.5-level performance on MMLU fell from $20 per million tokens in November 2022 to $0.07 by October 2024 — a reduction of more than 280 times.
The same report points to roughly 30% annual improvement in AI hardware price-performance. It also shows how much capability has migrated into smaller models. In 2022, the smallest model exceeding 60% on MMLU was Google’s PaLM at 540 billion parameters. By 2024, Microsoft’s Phi-3-mini crossed the same threshold with 3.8 billion parameters: a 142-fold reduction in model size.
None of that means every form of intelligence will inevitably become 280 times cheaper every two years, or that smaller models will always match larger ones. Benchmarks are imperfect proxies for usefulness and different workloads have very different economics.
But the direction is difficult to ignore.
Useful intelligence is becoming cheaper to produce, cheaper to access and available in more forms.
That changes the economics of where AI creates value.
Frontier intelligence is being sold at different price points#
By late 2025, the frontier model market itself was beginning to make this increasingly visible.
OpenAI launched GPT-5 in August as a family spanning GPT-5, GPT-5 mini and GPT-5 nano. In the API, input pricing at launch ranged from $1.25 per million tokens for GPT-5 to $0.05 for GPT-5 nano. The point is not that those models offered identical capability. They did not.
The interesting thing is that developers could increasingly choose how much intelligence, speed and cost a particular task justified.
Google was doing something similar with Gemini. Gemini 2.5 Pro and Flash became generally available in June, alongside the introduction of Flash-Lite as the fastest and most cost-efficient member of the 2.5 family. Google explicitly positioned the models around different trade-offs between reasoning performance, cost and speed.
Anthropic was pushing in another direction. Claude Sonnet 4.5, launched in September, extended Claude’s strength in coding, computer use and complex agents while retaining Sonnet 4’s API pricing. As I argued in Claude Is Becoming More Than a Chatbot, frontier intelligence is also becoming more capable of acting rather than merely producing an answer.
These developments are different, but economically they point in a similar direction.
Businesses are gaining access to a widening spectrum of machine intelligence: different capabilities, different latencies, different degrees of autonomy and different prices.
The question gradually changes from Can we afford intelligence for this task? to What level of intelligence does this task actually require?
That is a very different constraint.
Abundant intelligence creates abundant possibilities#
Imagine that the cost of producing a reasonably good analysis, idea, prototype, forecast, recommendation or piece of software falls dramatically.
A company does not merely get the same amount of work for less money.
It can explore things that previously would not have justified the cost.
Ten strategic scenarios can become a hundred. Three product concepts can become thirty. One analysis can be rerun across every customer segment. Software ideas that once required a development sprint can be prototyped before anyone decides whether they deserve one.
Agents potentially extend this further because intelligence can increasingly participate in sequences of work rather than isolated outputs.
The result is not a shortage of possibilities.
It is an explosion of them.
And that creates a new problem.
When ideas are expensive, generating more ideas creates value. When ideas become abundant, choosing between them creates value.
A business cannot pursue every plausible opportunity simply because AI has made it cheaper to identify or explore. The argument in You Can’t Pursue Every AI Opportunity becomes more important as the cost of generating opportunities falls.
AI can lower the cost of asking What could we do?
It does not remove the need to answer What should we do?
Intelligence and judgement are not the same thing#
This distinction matters because we often use the word intelligence to cover several different activities.
A model might be able to analyse a market, generate strategic options, identify patterns in customer data, critique a plan and estimate possible outcomes.
Those are valuable capabilities.
But a business still has to decide which objective matters most, which trade-offs it is willing to make, which risks are acceptable, which customers it wants to serve and which opportunities deserve scarce organisational attention.
Those decisions depend on more than analytical capability.
They depend on judgement.
Intelligence helps determine what could be done. Judgement determines what is worth doing.
That judgement can include experience, values, incentives, timing, appetite for risk and an understanding of consequences that are difficult to encode fully in a prompt.
The better AI becomes at generating plausible options, the more important the ability to distinguish between a plausible option and the right option may become.
General intelligence makes specific knowledge more important#
There is another consequence of widely available frontier models.
If your competitor can access broadly the same model you can, access to that model is unlikely to remain a durable advantage by itself.
The difference increasingly lies in what surrounds it.
What does the system know about your customers that a general model does not? What does it understand about your economics, products, operational constraints, history, processes and risk? What proprietary data can it reason over? What organisational knowledge has been made accessible to it?
This is why your AI doesn’t know your business unless you deliberately give it the context required to do useful work.
As models become more generally capable, the differentiated layer may increasingly be the specific context through which that capability is applied.
As general intelligence becomes less scarce, specific knowledge may become more valuable.
Two companies may use the same frontier model and receive very different economic value from it because one has connected the model to rich organisational context and the other has not.
The model is shared infrastructure.
The context is not.
Competitive advantage moves up the stack#
This suggests an important change in how businesses should think about AI strategy.
There will still be moments when access to a particular model or technical capability provides an advantage. Frontier capability does not diffuse instantly, prices are not identical and implementation quality varies enormously.
But advantages based purely on access are vulnerable when the underlying technology is improving rapidly and being sold to everyone.
Competitive advantage therefore moves towards things that are harder to commoditise.
Knowing which business problem is genuinely valuable. Recognising an opportunity before others do. Possessing proprietary context. Redesigning a workflow around a new capability. Getting employees to use it effectively. Understanding when a model’s recommendation should be trusted and when it should be challenged. Converting saved time into actual economic capacity. Choosing not to pursue an attractive-looking idea because something else matters more.
None of these are solved simply by buying a better model.
In fact, more capable models can make some of them harder because they increase the number of credible things a company could attempt.
The frontier can expand faster than an organisation’s ability to decide what to do with it.
The bottleneck becomes deciding where intelligence matters#
For years, organisations have rationed high-quality thinking because high-quality thinking was expensive.
AI begins to relax that constraint.
But organisations still have finite capital, finite management attention, finite implementation capacity and finite tolerance for change.
So the economic bottleneck moves.
The limiting factor is less likely to be whether a company can generate another analysis, another idea or another possible use of AI. It is increasingly whether the organisation can identify the opportunities that matter, supply the context needed to pursue them and exercise the judgement required to turn capability into value.
This is why chasing whichever frontier model happens to lead a benchmark is not an AI strategy.
The technology matters enormously. But as the technology becomes more capable and more accessible, the strategic question becomes less about possessing intelligence and more about directing it.
AI doesn’t eliminate scarcity. It moves it.
And if useful intelligence continues becoming cheaper and more abundant, the scarce resource may increasingly be the human and organisational judgement that determines where all that intelligence should go.