The Cost of Intelligence Is Collapsing
What a 214x fall in inference prices tells us about agents, work, and the next economy
I’ve been saying we have been inside the singularity since 2022 and the evidence of this is building.
The magnitude of this change and the accelerating speed of it is shocking everyone in data science.
As I’ve mentioned in the past, for most of human history intelligence was expensive because it came packaged inside a person.
If you wanted more analysis, you hired an analyst. If you wanted more code, you hired an engineer. If you wanted more research, planning, design, writing, or judgment, you found people with the right skills and paid for their time. Software made those people more productive, but the person remained the scarce unit at the center of the work.
That arrangement is starting to break.
Epoch AI has assembled one of the clearest measurements of the change. Its inference price analysis asks a simple question: what was the lowest API price required to reach a fixed level of model performance at different points in time?
For a GPT-4-level threshold on MMLU, the answer fell from $37.50 per million weighted tokens in March 2023 to $0.175 in February 2025.
That is a 214× decline in less than two years.
The number matters because it holds capability roughly constant. It is not saying that the same model became cheaper. It is saying that the market learned how to deliver a comparable benchmark level of intelligence for dramatically less money.
That is the curve we need to understand.
The AI conversation still focuses mostly on which model is smartest. But the larger economic question is how quickly adequate intelligence is becoming cheap enough to use everywhere. Once capable inference becomes inexpensive, we stop rationing it. We put it inside every product, every workflow, every decision, and every machine.
The cost curve bends down at the same time the demand curve bends up.
How the 214x Analysis Works
The logic is simple. Epoch combines public model benchmark scores with API pricing data. It sets a capability threshold, then identifies the lowest-priced non-reasoning model available at each point in time that meets or exceeds that threshold.
For the series we are looking at, the threshold is an MMLU score of 86, roughly the level reached by the original GPT-4 release.
MMLU is a broad general-knowledge benchmark. It is not intelligence in full. No benchmark is. But it gives us a consistent measuring stick across model generations.
The progression is striking:
GPT-4 launched at a weighted price of $37.50 per million tokens.
GPT-4 Turbo brought that down to $15.
GPT-4o brought it to $7.50.
Gemini 1.5 Pro reached the threshold at $2.19.
Gemini 2.0 Flash reached it at $0.175.
Epoch calculates the token price as a 3:1 weighted average of input and output prices. First-party pricing is used when available; otherwise, the analysis uses the median across API providers. Reasoning modes are excluded because they can generate far more internal and output tokens, making a simple price-per-token comparison misleading.
Epoch then repeats the process across multiple benchmarks and capability thresholds. The measured price declines vary widely, from 9× to 900× per year, with a median of 50× per year. The underlying notebooks and data are public, which makes the result more useful than another floating chart with no visible method.
There are important limits.
A token is not a completed task.
MMLU does not measure every form of reasoning, reliability, tool use, latency, context management, or domain expertise.
API list price is not the same as the provider’s underlying economic cost.
A model that meets the threshold on a benchmark may still behave very differently inside a real workflow.
The analysis does not prove that all intelligence became 214 times cheaper. It shows that one useful, fixed level of machine capability became available at 1/214th of its prior listed token price.
That is enough to change behavior.
The Price of Adequate Intelligence
The frontier gets the headlines, but in reality, the economy is transformed by what happens several steps behind it.
Most work does not require the smartest model in the world on every call. It requires a model that is good enough for the task, connected to the right context, given the right tools, bounded by clear instructions, and checked by an evaluation loop.
This is the idea of adequate intelligence.
An accounts-payable workflow does not need to solve open problems in physics. A sales-research agent does not need to write a great novel. A code-review agent does not need to manage a board meeting. Each system needs enough cognition to perform a bounded job reliably.
When that level of cognition is expensive, we use it sparingly. We reserve it for high-value queries and keep humans inside every intermediate step. When it becomes cheap, we redesign the workflow around its availability.
We let the agent read every document instead of a sample.
We let it test ten approaches instead of drafting one.
We let it monitor continuously instead of checking weekly.
We let it retry, compare, criticize, reconcile, and escalate.
This is not just a cheaper answer. This is a larger amount of cognition applied to the problem.
That distinction is where the economic implications begin.
Why Chat Becomes Command
In The Shift From Chat to Command, I called out that the interface to AI is changing from conversation to delegation.
A chatbot produces a response. An agent produces an outcome.
The prompt becomes the work order. The thread becomes the workspace. The human moves from performing each task to defining the objective, supplying context, granting tools, supervising execution, and deciding what ships.
But command is inference-intensive.
A chat answer may require one model call. An agentic workstream may require hundreds or thousands. The agent reads files, searches sources, calls tools, writes artifacts, runs tests, checks results, repairs failures, and asks for approval. A multi-agent system may explore several paths in parallel and discard most of them.
None of that works economically if capable inference remains precious.
The move from $37.50 to $0.175 changes the feasible “design space”. A workflow that would have been irresponsible to run continuously at the old price can become infrastructure at the new one. The system can spend more tokens thinking, checking, and trying again while still producing an outcome that is cheap relative to human labor.
This is why the shift from chat to command is not just a product-design choice. It is being pulled forward by the intelligence cost curve.
Cheap answers create chatbots while cheap sustained cognition creates agents.
Why Agents Become an Energy Problem
In Agents Per Gigawatt, I argued that the future of productivity will not be measured only as output per person. It will also be measured by how much useful synthetic labor a person, company, or country can command.
If agents are labor, inference is the fuel. Compute consumes energy. Energy flows through data centers. Data centers become factories for cognitive work.
The intelligence cost curve strengthens that thesis because it accelerates demand for inference. As the dollar price of adequate cognition falls, more use cases clear the economic threshold. More agents run. They run longer. They work in parallel. They monitor systems that were previously checked by people only when something went wrong.
This is where the apparent contradiction resolves.
Intelligence can become cheaper while the physical infrastructure behind it becomes more valuable.
The token price falls. Total token use explodes. Model efficiency improves. Aggregate compute demand rises. Data centers consume more power even as each unit of useful cognition costs less.
This is classic abundance behavior. When the cost of a useful input collapses, we rarely hold consumption constant and pocket the savings. We discover new ways to spend it.
Cheaper lighting did not make societies want less light. Cheaper bandwidth did not make us transfer less data. Cheaper compute did not make us run fewer calculations.
Cheaper intelligence will not make us think less.
It will create ambient cognition.
That is why power generation, grid access, chips, cooling, land, fiber, transformers, and data-center efficiency become part of the labor stack. The more affordable intelligence becomes in dollar terms, the faster demand runs into physical constraints.
Walking the Curve Forward
We should be careful here. Exponential curves make fools of anyone who turns them into a precise point forecast.
The right way to walk the intelligence cost curve forward is to separate three different curves that are often collapsed into one.
The first is price per token at a fixed capability level.
The second is cost per completed task.
The third is cost per reliable outcome inside a real operating system.
The first curve is falling fastest. Model compression, better architectures, improved hardware, higher utilization, batching, competition, and lower provider margins can all reduce the listed price of a token. Epoch explicitly notes that it did not attempt to isolate how much each factor contributed.
The second curve falls more slowly because tasks expand to consume available intelligence. A better agent may read more context, perform more tool calls, run more checks, and explore more alternatives. Tokens get cheaper, but the workflow deliberately uses more of them.
The third curve is the one businesses ultimately care about. It includes reliability, supervision, integration, security, failure recovery, and the cost of the surrounding harness. A cheap model inside a weak system can still be expensive because humans must repair the output. A more expensive model inside a disciplined workflow can be cheaper because it finishes the work correctly.
This is why price per token will eventually become a background metric, like the price of a floating-point operation. Important to infrastructure planners. Nearly invisible to the end user.
The useful unit will become cost per accepted outcome.
Still, the token curve shows the pressure underneath the system. Using the study’s February 2025 GPT-4-level observation of $0.175 per million weighted tokens as a starting point, even a much slower 3x annual decline would bring the price to roughly 1.9 cents by 2027 and less than one-tenth of a cent by 2030.
A 10x annual decline would make the raw token price economically trivial far sooner.
Those are scenarios, not forecasts. They are also the point.
If the direction persists at even a fraction of the historical rate, the economy stops asking whether it can afford a model call. It starts asking how much cognition should be applied to each objective, how many parallel attempts should run, and what system will judge the results.
The constraint moves from generation to selection.
The Demand Explosion
The first-order effect of cheaper intelligence is cost savings. The second-order effect is demand creation. The second-order effect will be much larger.
Today, companies ration analysis because people are expensive and attention is finite. Customer research is periodic. Compliance review is sampled. Forecasts are updated on a schedule. Software tests cover the paths the team had time to write. Managers accept incomplete information because gathering more is not worth the labor.
When cognition becomes cheap, the default changes.
Most of the technologists I know along with the broader futurist and “future of work” community believes (with good evidence) we will see these soon:
Every customer can receive individualized support.
Every contract can be reviewed.
Every product decision can be challenged by competing models.
Every code change can be tested against a growing library of failure cases.
Every operational system can be watched continuously.
Every executive can receive a live synthesis of the company instead of waiting for the monthly deck.
The geeky way to put this is the “long tail becomes economically reachable” but I think it’s going to feel like a never-ending renaissance of new experiences and improved living.
Markets that were too small to support human specialization can support agentic service. Languages that were too costly to localize can receive native-quality interfaces. Diseases with too few patients to attract large research teams can receive persistent machine attention. Small businesses can access analysis, design, software, finance, and operations capacity that previously belonged to large enterprises.
The cost decline makes previously uneconomic work exist.
This is why aggregate inference spending can rise even as the unit price collapses. The world will consume vastly more cognition because the number of useful places to apply it is close to unlimited already, and we’ll use our rising intelligence to find + create more surface area.
Human Capital Gets Repriced
When execution is expensive, the person who can produce the artifact holds leverage. When execution becomes cheap, leverage moves to the person who can define, direct, verify, and integrate the work.
This does not mean human knowledge becomes worthless. It means the value moves up the stack. Domain knowledge is what lets an operator specify the real objective, notice the hidden constraint, reject a plausible error, and understand which tradeoff the model cannot make on its own.
The analyst of the next decade does not compete with one agent. The analyst competes with another analyst who can direct fifty agents, compare their work, and find the signal inside the volume.
The manager does not win by protecting a larger headcount. The manager wins by building a system that turns a small amount of human judgment into a large amount of reliable execution.
The company does not win because it bought access to the best model. Model access diffuses. The company wins because it has captured context, mapped workflows, clear permissions, proprietary data, strong evaluations, trusted decision rights, and people who know what good looks like.
The model is the engine.
The harness is the productive system.
As intelligence gets cheaper, weak harnesses become more dangerous. A system that can generate one bad answer is annoying. A system that can generate, distribute, and act on a million bad answers is an operational failure.
Abundance punishes weak judgment.
The organizations that understand this will invest in evaluation and control at the same time they invest in generation. They will not measure adoption by the number of employees with chatbot accounts. They will measure it by reliable workstreams, accepted outcomes, cycle-time compression, and the ratio of machine execution to human review.
The Bottleneck Migration
The intelligence cost curve is one of the clearest empirical signs that scarcity is moving.
First, execution becomes abundant. Models can draft, analyze, code, classify, compare, and plan at a cost that keeps falling.
Then demand concentrates on the physical layer: compute, chips, electricity, grid access, cooling, and capital.
Then it moves again into the sovereign layer: trusted data, secure systems, institutional competence, decision rights, legitimacy, judgment, and the ability to turn machine output into real-world action.
For an individual, the bottleneck shifts from hours available to clarity of intent and capacity to review.
For a company, it shifts from hiring enough people to encoding enough context and building enough trust in the workflow.
For a country, it shifts from population alone to energy systems, compute infrastructure, capital formation, data access, security, and institutions capable of deploying synthetic labor without losing control of it.
This is the Sovereign Singularity thesis in economic form. Digital intelligence becomes abundant. Physical capacity and human agency become more valuable because they determine how abundance is aimed.
The winners will not control intelligence. They will command it.
What to Build Now
The practical response is not to wait for the curve to finish. Curves like this never finish. They change the design assumptions underneath the work while most organizations are still budgeting around last year’s costs.
Start with six moves.
Measure outcomes, not prompts. Track the cost of completed, accepted work. Token price is an input. The result is the product.
Build evaluation before volume. Cheap generation without reliable selection creates cheap slop at industrial scale.
Capture context. Write the operating procedures, connect the source systems, define permissions, and preserve the institutional knowledge agents need to act.
Route intelligence. Use the cheapest model that can perform each step, then escalate difficult cases to stronger models and humans.
Spend the savings on search. Run competing approaches, test assumptions, explore edge cases, and use abundance to widen the solution space.
Protect the final mile. Keep human accountability where trust, taste, ethics, relationships, and irreversible decisions matter.
The goal is not maximum automation. The goal is maximum useful agency.
The best operators will use cheap intelligence to expand what they can attempt while becoming more deliberate about what they approve. They will not try to personally touch every intermediate step.
They will design the system, set the standard, review the exceptions, and decide where the machine workforce should go next.
Intelligence Becomes an Input
The 214x decline does not tell us when artificial general intelligence arrives. It does not prove that benchmarks equal human judgment. It doesn’t give us a clean forecast for the price of every task.
It tells us something more immediate.
A useful level of machine cognition is moving from scarce capability to cheap input.
That transition is enough to reorganize work. It is enough to turn chat into command, software into labor systems, data centers into workforce infrastructure, and energy policy into economic strategy. It is enough to let small teams command output that once required large institutions. It is enough to make evaluation, judgment, context, and trust the new scarce assets.
We spent the first years of the AI era asking how smart the models could become.
The next phase will be defined by how cheaply adequate intelligence can be deployed, how much of it we can run, and how well we can aim it.
The cost of intelligence is collapsing. The value of judgment is compounding.
That is the curve to watch.
Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!
Thank you for helping us accelerate Life in the Singularity by sharing.
I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.
Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.
To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏
