Syntic
All posts

Published October 4, 2026 in Perspectives

The Cost of Intelligence Is Falling. What Happens Next?

By Syntic AI

In early 2025, Epoch AI estimates, it cost about 30 cents a question to reach a 75% score on GPQA Diamond, a test of graduate-level science questions. Less than 18 months later, a newer model matched that score for about four hundredths of a cent. That is a 725-fold fall in the price of one fixed slice of capability.

Over the same period, spending moved the other way. According to the International Energy Agency (IEA), capital expenditure by the largest technology companies exceeded $400 billion in 2025, leading hyperscalers and newer cloud providers have announced about $715 billion for 2026, and electricity use in AI-focused data centres rose 50% in 2025.

The tension between those facts is the real story. The price of a given level of machine intelligence is collapsing. The total bill for intelligence is not, because cheaper capability creates more uses, the most valuable uses consume far more of it per task, and much of what surrounds a model does not follow its cost curve at all. Anyone pricing a product or budgeting an AI programme needs to keep both curves in view.

How to measure the price of intelligence

The careful studies fix a capability level and track what it costs to reach it over time.

Epoch AI did this in March 2025 across six benchmarks over three years. The price of reaching GPT-4's level on PhD-level science questions fell by about 40x per year. Across all the thresholds it measured, the rate ranged from 9x to 900x per year, and Epoch noted that the fastest drops were the most recent, so it was "less clear that those will persist."

The Stanford AI Index 2025 used the same logic for a more familiar milestone. A model scoring at GPT-3.5's level on MMLU cost $20 per million tokens in November 2022 and $0.07 by October 2024, a fall of more than 280-fold.

Later work points the same way with lower rates. Hans Gundlach, Neil Thompson and colleagues built a larger dataset from Artificial Analysis and Epoch data and found that the price of a given benchmark score fell 5x to 10x per year for frontier models across knowledge, reasoning, maths and software engineering. Epoch's September 2026 update, using a different method across five benchmarks, puts the decline at about 47% per quarter, or roughly 13x per year.

The estimates disagree on the rate and share the same caveats. They measure list prices, which include margins and possibly subsidies, not underlying costs. Benchmarks are imperfect proxies for useful work, and buyers rarely switch to the cheapest model that meets their bar. The defensible summary is that the price of a fixed level of capability has been falling by around an order of magnitude a year, with wide error bars and no guarantee that the pace holds.

Four forces behind the curve

Hardware. The AI Index reports hardware costs falling about 30% a year and energy efficiency improving about 40% a year. That explains only a small part of a 10x annual decline.

Algorithms. Epoch researchers estimated in 2024 that the compute needed to reach a given language-model performance level halved roughly every eight months between 2012 and 2023, with a 95% confidence interval of 5 to 14 months. Gundlach and colleagues, after separating out hardware gains and using open models to control for competition, estimate algorithmic efficiency progress at about 3x per year.

Smaller models trained by bigger ones. The cheapest points on these curves are small models. Both the AI Index and Epoch figures end with models such as Gemini 1.5 Flash-8B and Microsoft's Phi-4. Distillation, training a small model on the outputs of a larger one, is a large part of how that happens. When DeepSeek released its R1 reasoning model in January 2025, it also released six smaller models, from 1.5 billion to 70 billion parameters, distilled from it.

Competition. Providers sell into a market with close substitutes, including open-weight models anyone can host; the AI Index found the gap between leading open and closed models narrowed from 8% to 1.7% on one leaderboard in a year. Prices that fall partly because of competition can stop falling if competition thins, so the curve is a trend, not a law.

Why the bill rises anyway

The same Gundlach paper contains the counterweight. While the price of a fixed score fell 5x to 10x a year, the price of running frontier models rose by 3x to 18x a year, because models grew larger and reasoning demanded more tokens.

Reasoning models make this visible. OpenAI's documentation says its reasoning models may generate "anywhere from a few hundred to tens of thousands of reasoning tokens" per problem, depending on complexity, and that these hidden tokens are "billed as output tokens." A cheaper token does not mean a cheaper answer if the answer takes a hundred times as many tokens.

Agents multiply the effect. Anthropic reported in June 2025 that in its data, agents "typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens" than chats. Its conclusion was economic: such systems need tasks "where the value of the task is high enough to pay for the increased performance."

The IEA sees the same pattern in energy: energy use per AI task has dropped by at least an order of magnitude a year, while video generation, reasoning and agentic tasks "can consume hundreds or thousands of times more energy per query than simple text generation."

So the market is splitting. A fixed capability, such as classifying a ticket, gets cheaper every quarter. A frontier task, such as a long agentic run on a codebase, may cost more each year, because buyers keep asking for more of it.

Jevons, with the right caveat

In 1865 William Stanley Jevons argued, about coal, that "it is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth." More efficient engines made coal worth using in more places, so total use rose.

The idea is often applied to AI too casually. In energy, the evidence for full "backfire", where efficiency gains increase total use, is weak: a 2016 review by Kenneth Gillingham, David Rapson and Gernot Wagner found "little support" for it. Rebound is real but usually partial. Whether backfire happens depends on how strongly demand responds to price.

For AI inference, demand currently looks very responsive. Alphabet said in July 2026 that its model APIs were processing about 22 billion tokens per minute, up from 16 billion a quarter earlier. The IEA found global data centre electricity demand grew 17% in 2025, and it projects data centre consumption roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, around 3% of global electricity demand. The IEA names the mechanism directly: efficiency gains are "fuelling broader adoption of AI services, a dynamic known as the Jevons paradox."

That projection is uncertain. The IEA notes that investment has outgrown company balance sheets, depends on capital markets, and faces shortages of grid connections and high-bandwidth memory. Demand can be elastic and still hit physical limits.

As with customer service when every call can be answered, lowering the cost of access reveals demand the old price was suppressing.

What becomes newly economical

The useful question for a business is not "what will be cheaper?" but "what was never done because it cost too much?"

Coverage is the clearest category. Many processes sample because full review is unaffordable: some contracts checked for a clause, some support conversations audited. When reading a document costs fractions of a cent, sampling becomes a choice rather than a constraint.

Redundancy is another. Cheap capability can be spent on reliability: running several attempts and selecting the best, adding a second model to check the first, or verifying an output against a source before it reaches a person. The economics of using more than one model depend on exactly this spread between cheap and expensive capability.

Then there are markets that were too small to serve. Long-tail languages, low-value transactions and customers who never justified a person's time become serviceable once the marginal cost of an interaction falls below its value.

What does not get cheaper

Philippe Aghion, Benjamin Jones and Charles Jones, writing on AI and growth in 2017, drew on William Baumol's "cost disease" to argue that growth "may be constrained not by what we are good at but rather by what is essential and yet hard to improve." The same applies inside an AI deployment. As inference falls in price, the expensive parts become everything else.

Human verification. A reviewer's hour costs what it cost last year. If output volume rises tenfold and review stays human, review becomes the main cost and the main bottleneck.

Data. Clean, permissioned, current data about a company's own customers, products and processes is not on any price curve. It has to be collected, governed and maintained.

Integration. Connecting a model to systems of record and keeping workflows running through change is engineering work, priced in engineers.

Liability. The EU's revised Product Liability Directive (2024/2853) brings software, including AI systems, within strict product liability, and member states must transpose it by 9 December 2026. Cheaper intelligence does not lower the cost of a wrong answer. It raises the number of answers that might be wrong.

Physical inputs. Epoch's September 2026 analysis notes that the AI boom is strong enough to raise prices for the inputs it consumes: "chips, power, even the labor of electricians." The IEA expects the high-bandwidth memory shortage to last until at least the end of 2027.

Pricing software when marginal cost swings

Traditional software pricing assumed near-zero marginal cost: a seat cost the vendor about the same whether the customer used it once a month or all day. AI breaks that assumption, and the break is visible in public pricing changes.

In June 2025, GitHub began enforcing monthly allowances of "premium requests" for Copilot, with multipliers for more expensive models. On 1 June 2026 it replaced request-based billing with usage-based billing, where cost depends on "the model and the number of tokens consumed," drawn from an allowance of AI credits.

Cursor made a similar move in mid-2025, shifting its Pro plan from a count of requests to $20 of included usage. Its explanation was blunt: "new models can spend more tokens per request on longer-horizon tasks," and "the hardest requests cost an order of magnitude more than simple ones." It also apologised for how the change was communicated and offered refunds.

Whatever their effect on individual bills, these changes reflect what happens when the cost of serving one customer action can vary by orders of magnitude. Expect more products to price in credits, allowances and outcomes, with per-seat pricing surviving mainly where usage is predictable. For buyers, the line to watch is how heavy, reasoning-intensive work is metered. That shift is part of why intelligence is becoming infrastructure, sold more like a utility than a licence.

Planning against two curves

A sensible AI plan tracks two numbers.

The first is the price of the capability you need today. If it keeps falling by around an order of magnitude a year, designs that are marginal now may be cheap within a budget cycle, and multi-year contracts at today's unit prices deserve scrutiny.

The second is the cost of the capability your users will want next. Longer tasks, more reasoning and more autonomy pull that number up, and it may rise even as the first one falls.

The errors are symmetrical. Treat all AI as cheap and reasoning-heavy workloads will overrun the budget. Treat all AI as expensive and competitors will build on the cheap layer first. The organisations that get the economics right will know which of their workloads sits on which curve, and will spend the savings from the first on the people, data and checks that the second still requires.

Sources