The Real Cost of AI: What Does It Cost to Keep a Machine Intelligent?
AI models are becoming dramatically cheaper to use. But deployment is only the beginning. The bigger question for businesses may be what it costs to keep artificial intelligence useful once the world it learned from begins to change.
Artificial intelligence has acquired a strange economic property. It is simultaneously attracting extraordinary amounts of money and becoming remarkably cheap to use. The decline in inference costs alone has been dramatic. Stanford University’s AI Index found that the cost of querying a model performing at roughly GPT-3.5 level on the MMLU benchmark fell from $20 per million tokens in November 2022 to just $0.07 by October 2024, a reduction of more than 280-fold in around 18 months. Depending on the task measured, Stanford found annual declines in inference prices ranging from ninefold to 900-fold.
The pressure has continued. In July 2026, OpenAI cut the price of GPT-5.6 Luna by 80 per cent, bringing standard short-context API pricing to $0.20 per million input tokens and $1.20 per million output tokens. Anthropic has meanwhile made the introductory $2 input and $10 output price per million tokens for Claude Sonnet 5 permanent, abandoning a previously planned increase. The models are not directly equivalent, and headline token prices tell us little about how efficiently either will complete a particular task, but the direction of competition is unmistakable: intelligence is being sold more cheaply.
At the same time, spending on AI is going the other way. Stanford estimates that global corporate AI investment reached $581.7 billion in 2025, an increase of 130 per cent in a year, while global AI compute capacity has been expanding rapidly to accommodate the demand for training and inference.
There is no real contradiction here. The apparent paradox exists because the price of an AI model and the cost of artificial intelligence are not the same thing. For a business, the real cost of AI extends beyond subscriptions and tokens to encompass infrastructure, data, integration, monitoring, human oversight, evaluation, governance, retraining and the consequences of poor decisions. A model can become considerably cheaper to query without the system built around it becoming correspondingly cheaper to own.
Once AI starts influencing decisions rather than merely generating text, another cost enters the calculation. It is harder to measure and rarely appears on a technology invoice, yet it may ultimately matter far more than the price of inference: the cost of an AI system continuing to operate successfully from a technical point of view while becoming progressively less accurate about the world in which it is operating.
The interesting question, then, is no longer simply how cheaply we can create intelligence. It is what it costs to keep that intelligence useful.
The cheapest part of AI is getting cheaper
API pricing is seductive because it is easy to understand. A provider publishes a price for input and output tokens, a company estimates its consumption, and a relatively neat number emerges. For straightforward applications, that calculation may provide a reasonable approximation of cost. If a business wants a model to summarise a known volume of documents or classify a predictable stream of customer enquiries, model usage may indeed be one of the clearest items on the bill.
The calculation becomes less satisfactory once AI is embedded in the machinery of a business. Data needs to reach the system in the correct form; outputs need to connect to existing software; permissions, security and access have to be managed; performance needs to be evaluated; failures have to be identified; and someone has to decide what happens when the machine's recommendation conflicts with human judgement. At this point AI begins to look less like a software subscription and more like an operating system sitting inside another operating system: the business itself.
This matters particularly in applications where an AI system is attempting to predict or optimise something that changes over time. Think of a model predicting equipment failure in a factory, forecasting electricity demand, identifying fraud, allocating inventory between warehouses or adjusting decisions in an energy market. The economic value of these systems does not lie in producing large numbers of inexpensive predictions; it lies in those predictions continuing to lead to better decisions.
The difference is subtle but important. The cost per prediction can fall while the cost per useful prediction rises.
Deployment is not the finish line
Machine learning introduces a maintenance problem that conventional software does not encounter in quite the same way. A piece of software usually continues to perform the same calculation until the software itself, its requirements or the systems surrounding it change. A machine-learning model can remain technically intact while the relationship between the information it receives and the world it is trying to understand changes around it.
The US National Institute of Standards and Technology treats this as an important part of operating AI responsibly. Its AI Risk Management Framework recommends ongoing monitoring because model performance and trustworthiness can change after deployment, including through performance degradation, unexpected behaviour and changing conditions.
The underlying problem is less complicated than the terminology surrounding it suggests. Imagine that a retailer trains an AI system to forecast demand using several years of relationships between prices, seasons, weather, promotions, location and customer behaviour. The software does not suddenly stop running when those relationships change. Perhaps customers begin buying differently, delivery patterns alter, a competitor enters the market or a product that once sold predictably during warm weather suddenly falls out of fashion. The system will still accept the data and return a forecast, but the world represented by its training may gradually cease to resemble the world in front of it.
There are distinctions within this. Data drift broadly refers to changes in the statistical characteristics of the information an AI system encounters, while concept drift concerns changes in the relationship between that information and whatever the model is attempting to predict. Not every shift automatically causes a model to deteriorate, which is precisely why monitoring matters: organisations need to know whether change in the environment is actually changing the quality of the decisions being made. NIST accordingly recommends comparing production performance with pre-deployment results and establishing processes for detecting and responding to deterioration.
The uncomfortable consequence is that an AI system can be functioning precisely as designed while becoming steadily less valuable.
The hidden cost of AI is often the cost of being wrong
This is where conventional calculations of AI expenditure begin to break down. Suppose a manufacturing company uses a model for predictive maintenance and inference prices fall sufficiently for its annual model bill to drop by 40 per cent. On the technology budget, AI has become cheaper. Yet the machinery being monitored has aged, some components have been replaced, the factory is being operated differently and production patterns have changed. If the relationships learnt by the original model no longer describe those conditions accurately, the economic consequences of deteriorating predictions could overwhelm the saving made on compute.
Research into continual learning for predictive maintenance describes exactly this problem. Conventional deep-learning models are typically trained against a fixed distribution representing conditions at a particular point in time, while industrial environments evolve. Changes in equipment, operating contexts and external conditions can therefore reduce performance when a model lacks the ability to generalise or adapt.
If poorer predictions lead engineers to replace components unnecessarily, miss an emerging failure or operate machinery inefficiently, the important cost is no longer an API charge. The same principle extends across the economy. A fraud system that is slow to recognise a new pattern of behaviour can create financial losses; a demand model that persistently overestimates a product leaves capital sitting in stock; an optimisation system based on old assumptions can waste energy; and a scheduling model that becomes only marginally less efficient can generate thousands of small operational losses over time.
These costs are particularly difficult to see because few of them will be recorded as expenditure on artificial intelligence. They are more likely to appear elsewhere in a company's accounts as maintenance, inventory, energy, labour, lost revenue, fraud or inefficiency. The model itself has a visible price, but the financial consequences of relying on a deteriorating model can be spread throughout the organisation.
Any meaningful calculation of the real cost of AI therefore needs to account not simply for what it costs to operate the system, but for what happens when the system's understanding of its environment begins to fall behind reality.
What does it cost to keep an AI model useful?
The conventional answer is some form of maintenance cycle. Performance is monitored, new data accumulates, engineers investigate whether the operating environment has changed sufficiently to matter, data is prepared and sometimes labelled, and the model is retrained or fine-tuned. The new version must then be evaluated against the existing one and checked against whatever accuracy, safety, robustness or regulatory requirements apply before it can replace the model already in production.
There is nothing inherently unsophisticated about this process. In many settings it is exactly what responsible machine learning should look like. The important point is simply that none of it is captured by the price per million tokens. NIST's guidance explicitly treats post-deployment monitoring and corrective action as part of the AI lifecycle, while the European Union's AI Act contains specific provisions for high-risk systems that continue learning after being placed on the market, as well as wider post-market monitoring requirements.
It is also worth separating two problems that are often blurred together: giving an AI access to current information and allowing the AI itself to learn. A large language model does not need to change its underlying parameters every time a new fact appears. Retrieval systems, databases, search engines and tools can provide live information without altering the model's weights. If an AI assistant needs today's exchange rate, the sensible answer is to connect it to an authoritative source for today's exchange rate rather than retrain it on the concept of currency.
The problem is different when the relationships governing the decision are changing. If machinery behaves differently as it ages, fraudsters alter their techniques, electricity markets move into a new regime or consumer behaviour develops in ways that the original model has never encountered, simply retrieving a fresh document may not be enough. The system may need to alter what it has learnt about the relationship between cause, context and outcome.
That process of adaptation has become one of the more consequential questions in applied AI.
Agents are making the old cost calculation even less useful
The rise of AI agents is already revealing how inadequate cost-per-token can become as a measure of economic value. A conventional chatbot interaction is relatively easy to imagine: a question is submitted, the model processes it and a response returns. An agent may interpret an objective, create a plan, search for information, call external tools, inspect what those tools return, reconsider its strategy, execute another action and continue through several cycles before completing the original request.
Google's current documentation for its managed Gemini agents says a single interaction can trigger multiple reasoning loops and typically consume between 100,000 and three million tokens. Particularly complex workflows involving many tool calls can consume even more.
That does not make agents economically unattractive. An agent costing several pounds to complete work that previously required hours of human labour could represent extraordinary value. What it demonstrates is that the unit price of the underlying model tells us very little in isolation. A cheap model that requires repeated attempts and human correction may be more expensive in practice than a more costly model that completes the task reliably in fewer steps.
This suggests a different metric for businesses. Instead of asking only about cost per token, they may increasingly need to understand cost per successful outcome. For AI systems embedded in long-running operations, there is a further extension of that idea: the total lifetime cost of producing reliable decisions.
That calculation has to include what happens after deployment.
What if learning did not end when the model went live?
Continuous or continual learning begins from a fairly simple observation: the boundary between training an AI system and using it may be too rigid for environments that never stop changing.
The terms surrounding the field are not entirely interchangeable. Continual learning generally describes systems that acquire knowledge sequentially while attempting to retain useful capabilities learnt previously. Online learning usually refers more specifically to algorithms that update incrementally as new observations arrive, while active learning describes systems that selectively determine which new examples would be most useful to learn from. Adaptive AI is a broader term that can encompass combinations of these approaches. What they share is a departure from the idea that an intelligent system should learn once, be deployed and remain essentially fixed until engineers decide to train another version.
The attraction is particularly obvious in environments where the rate of change is high. Instead of waiting until degradation becomes serious enough to justify a large retraining cycle, a system capable of continual adaptation could incorporate relevant new experience as it operates. In principle, this could shorten the gap between the world changing and the model responding to it, reduce the amount of historical data that needs to be processed repeatedly and make adaptation a routine part of operation rather than an occasional engineering project.
There is early research suggesting meaningful efficiency advantages are possible in particular circumstances. Work on continual learning under concept drift has shown that targeted adaptation can, on experimental benchmarks, achieve performance comparable with full relearning while using far less newly labelled data and computation. That result should not be extrapolated into a universal claim that continual learning is cheaper: it was produced under specific benchmark conditions and the field remains technically difficult. It does, however, illustrate why the economics are worth examining.
The potentially important saving may therefore be larger than compute alone. If an adaptive system can respond earlier to meaningful change, it could also reduce the length of time during which a business is making decisions using a model that has started to become stale.
In some applications, the value of continual learning may ultimately be measured in the mistakes that never happen.
The companies trying to close AI's learning loop
For years, continual learning has been primarily a research problem. What is changing is the appearance of companies attempting to turn different interpretations of the idea into commercial systems. They are not all building the same technology, and it would be misleading to group every adaptive product under a single definition, but their emergence suggests that the separation between training and deployment is beginning to be questioned outside academic laboratories.
Wakeline: continuous learning inside live decision systems
German AI company Wakeline approaches the problem from the perspective of decision-making in environments that continually produce new information. Its stated architecture places learning inside the live system rather than treating it as a separate retraining stage: the system observes signals, retains information about outcomes and updates its internal decision-making as conditions change. Wakeline's commercial Market Edge application applies the approach to battery energy storage and electricity markets, where the company says its model can update its assumptions as market behaviour evolves rather than relying on conventional manual retraining cycles.
What makes the approach relevant to the economics of AI is the distinction Wakeline is drawing between a model that is periodically replaced and an intelligence system that remains embedded in the environment from which it learns. Electricity markets are a useful test case precisely because they do not offer the luxury of permanence: prices, demand, renewable generation, storage behaviour and market regimes continually interact. Wakeline's proposition is that adapting to those outcomes should be part of the architecture of the intelligence itself rather than an external maintenance process. These remain company claims rather than evidence that continuous learning will eliminate retraining costs universally, but they point towards a fundamentally different way of thinking about AI lifecycle economics.
Trajectory: learning from what users actually do
A different interpretation is emerging in generative AI. Trajectory, launched in 2026 by researchers with backgrounds including Google DeepMind and Apple, describes itself explicitly as a platform for continual learning. Its system captures information generated when people actually use an AI product, corrections, retries, edits and other signals, and feeds that experience into continuing post-training and evaluation. The company raised $15 million in seed funding and has attracted backing from figures including Google DeepMind chief scientist Jeff Dean and Stanford professor Fei-Fei Li.
Trajectory is interesting because it treats product usage itself as training material. Rather than assuming an AI product reaches the end of its learning process when released, the company is attempting to build a loop in which real interactions reveal where the system succeeds and fails, new versions are trained against those signals, and improvements are passed through evaluation and approval before deployment. Its own platform emphasises that companies retain control over which interactions are used for training and that updates pass through evaluation before reaching production, an acknowledgement that continual learning requires governance as much as learning machinery.
This is not necessarily continuous learning in the sense of a model rewriting itself after every conversation. Indeed, Trajectory's founders have acknowledged that early implementations update models periodically rather than after each individual interaction. The important shift is architectural: real-world usage is no longer treated merely as telemetry to inspect after something goes wrong, but as part of an ongoing learning process.
Lightwheel: continuous learning reaches the physical world
The problem becomes even more tangible when AI controls machines rather than software. Lightwheel is building what it calls continuous-learning infrastructure for Physical AI, linking simulation, human-generated data, model evaluation and real-world robot deployment into a feedback loop. Its aim is to allow weaknesses identified after deployment to influence what data is collected and what scenarios are subsequently simulated and trained.
This reflects a problem robotics has faced for decades: the real world contains too much variation to capture perfectly before deployment. A robot trained in simulation will eventually encounter a grip, object, surface, movement or sequence that differs from what its designers anticipated. Lightwheel's approach is therefore based on a Real2Sim2Real loop in which real-world experience can identify gaps, simulation can generate new learning environments around those gaps, and updated policies can then be evaluated and returned to deployment. The company has described this process in work with industrial partners including Geely, although performance claims from such programmes should be read as company-reported results rather than independent benchmarks.
Physical AI makes the underlying argument unusually clear. If a warehouse, factory or hospital changes after a robot has been deployed, intelligence that cannot accommodate those changes will eventually become less useful. The question is how to allow adaptation without surrendering control over what the machine learns.
Active learning offers another route
Other companies are approaching part of the same economic problem without claiming full continual learning. Secondmind, for example, uses active learning in engineering applications to identify which experiments or data points would provide the most useful information next. Rather than continuously modifying a production model in response to every new observation, its technology is intended to reduce the amount of expensive engineering testing and data collection required to improve models.
That distinction matters. Active learning, online learning and continual learning should not be collapsed into a fashionable umbrella term, yet each represents an attempt to make learning more responsive and data-efficient than repeatedly starting from a large static dataset. The commercial landscape forming around these approaches suggests that the next phase of AI efficiency may concern not simply how cheaply a model can run, but how efficiently it can acquire the next piece of useful knowledge.
Why continuous learning is not an automatic solution
If continuous learning sounds like an obvious improvement, the difficulties explain why it is not already standard.
One of the best-known problems is catastrophic forgetting, in which learning new information damages knowledge or capabilities acquired previously. The broader stability-plasticity problem follows from the same tension: an adaptive system must be plastic enough to respond to meaningful change without becoming so plastic that new information constantly destabilises what already works. Continual-learning research therefore attempts to balance adaptation with memory rather than simply maximising the rate at which a model changes.
There is a second problem that is arguably more troubling for systems intended to learn for years. Research published in Nature in 2024 found that deep neural networks can suffer a loss of plasticity during prolonged continual training, progressively reducing their ability to learn new information. The study found serious plasticity loss in both supervised and reinforcement-learning settings, demonstrating that simply continuing conventional training indefinitely does not automatically produce a lifelong learner.
Then there is the question of what counts as a useful lesson. Incoming data can be anomalous, biased, malicious, temporarily unrepresentative or produced by other automated systems. In some cases an AI system may even be learning from consequences that its own earlier decisions helped create. A pricing system observes customer behaviour, but that behaviour partly reflects prices the system itself chose. A recommendation engine influences what people encounter and then learns from their subsequent behaviour. A fraud system operates against adversaries who actively change their behaviour in response to detection.
The design problem is therefore not merely how to build AI that keeps learning. It is how to build AI that can decide, or be governed so that humans can decide, what deserves to become part of what it knows.
That is why the more credible commercial approaches to continual learning tend to include evaluation, memory, thresholds, approval mechanisms or bounded learning loops. The objective is not an AI system that changes its mind constantly. It is an AI system capable of changing when reality provides good reason to do so, without casually discarding what it previously learnt.
Sometimes retraining will still be the better answer
Continuous learning should not become an ideology. There are many situations in which periodic retraining will remain simpler, safer and economically preferable. If the environment changes slowly, training data is easy to collect and the consequences of an incorrect adaptive update are severe, a controlled release cycle may be exactly the right design.
There are also many cases where updating the model is unnecessary. Connecting a general-purpose AI system to current, authoritative information may solve a knowledge problem far more safely than continually changing the model itself. Continual learning becomes most compelling where the underlying environment is genuinely non-stationary and where performance depends on continually understanding changing relationships, rather than simply knowing the latest facts.
Nor is the academic problem solved. Research evaluating continual-learning techniques under more realistic conditions has found that methods performing convincingly on controlled benchmarks can struggle considerably once exposed to the complexity of real-world environments. A 2024 study of continual learning for 3D semantic segmentation concluded that the tested methods remained significantly below the performance attainable through joint offline training, warning against assuming that laboratory progress translates straightforwardly into production systems.
The meaningful question is therefore not whether continual learning is superior to traditional machine learning. It is when the cost and complexity of adaptation become lower than the cost of periodically falling behind.
Different businesses will arrive at very different answers.
The AI metric businesses may eventually care about
The public AI conversation has spent several years concentrating on two numbers: capability benchmarks and token prices. Both are useful, but neither tells a company whether an AI system is economically worthwhile over its operating life.
Imagine two systems attempting to perform the same job. The first relies on a very inexpensive model but needs frequent human checking, deteriorates noticeably as conditions change and requires substantial retraining several times a year. The second costs more each time it runs but remains reliable for longer, requires fewer interventions and can adjust successfully to certain changes in its environment. There is no way to determine which system is cheaper from their API price sheets.
A more useful concept is something approaching the total cost of reliable intelligence. That means looking beyond the cost of generating an answer to the cost of keeping the answer good enough to act upon. Model access, compute, integration, data engineering, monitoring, human supervision, retraining, validation and governance belong in that calculation, but so do the consequences of getting decisions wrong while a model waits to be updated.
Continuous-learning systems become economically interesting in this framework because they offer the possibility of altering the maintenance side of the equation. They might reduce the amount of intervention required to keep a system relevant, improve the efficiency with which new data is incorporated or shorten the period between meaningful environmental change and model adaptation. Whether those gains outweigh the extra technical and governance complexity will depend on the application.
This is also where the economics of AI could begin to diverge sharply from the economics of conventional software. If intelligence increasingly becomes something that learns from the environment in which it operates, then its value will depend not only on what it knew when purchased but on what it is capable of becoming while owned.
Cheap intelligence is only the beginning
The fall in inference prices is one of the most consequential developments in modern computing. Making powerful models dramatically cheaper allows artificial intelligence to be used in places where the economics would previously have made little sense. As prices fall further, AI will inevitably be asked to perform more work, make more recommendations and participate in more decisions.
That very success makes the maintenance problem more important. An AI system used occasionally to draft marketing copy can be replaced with relatively little consequence. An AI system influencing energy trading, industrial machinery, logistics, fraud detection or autonomous software cannot be judged so casually. The cost of its intelligence includes the cost of keeping that intelligence aligned with the conditions in which it has to operate.
The first great AI competition was largely about capability: who could build the most powerful model. The next has increasingly become one of efficiency, as OpenAI, Anthropic, Google, Chinese laboratories and open-weight developers compete to deliver more capability at lower prices. A third competition may now be forming around something more difficult to capture on a benchmark: who can maintain useful intelligence for the lowest lifetime cost.
For stable environments, conventional models, careful monitoring and periodic retraining may continue to provide the best answer. For systems operating in conditions that change continuously, the companies now experimenting with continual, online and adaptive learning are asking a more fundamental question about how artificial intelligence ought to work.
The cheapest model will not necessarily produce the cheapest intelligence. Once AI begins making decisions about the real world, its true cost will be determined not simply by how little we pay each time it thinks, but by how long it remains worth listening to.
Frequently Asked Questions
What is the real cost of AI for a business?
The real cost of artificial intelligence includes much more than API charges or software subscriptions. Businesses may need to account for infrastructure, data preparation, integration, security, monitoring, human oversight, model evaluation, governance, retraining and maintenance. For AI systems that influence operational decisions, the cost of deteriorating performance can also be significant because inaccurate decisions may create losses elsewhere in the business rather than appearing as direct AI expenditure.
Is AI becoming cheaper?
The cost of accessing AI models has fallen dramatically. Stanford's AI Index found a more than 280-fold reduction between November 2022 and October 2024 in the cost of querying a model achieving approximately GPT-3.5-level performance on one widely used benchmark, and competition between model providers has continued to drive prices down. This does not mean every AI implementation is becoming cheaper, however, because inference is only one component of the total cost of operating an AI system.
What is model drift?
Model drift describes deterioration that can occur when the data or real-world relationships encountered by an AI system change after it has been trained. The model may continue operating normally from a technical perspective while its predictions become less accurate or useful. Monitoring for these changes is therefore an important part of operating machine-learning systems in environments that evolve over time.
Why do AI models need retraining?
Models may need retraining when the information or relationships represented by their original training data no longer adequately describe current conditions. New products, customers, machinery, market regimes, behaviours or risks can all change a model's operating environment. Retraining allows newer information to be incorporated, although it also creates costs around data preparation, evaluation, validation and deployment.
What is continuous learning in AI?
Continuous or continual learning refers broadly to approaches that allow AI systems to acquire new knowledge after their initial training while attempting to retain useful knowledge acquired previously. Rather than treating learning as something that ends before deployment, continual-learning systems attempt to make adaptation part of an AI system's operating lifecycle. Continual learning is related to online and active learning, but the terms describe different techniques and should not be treated as exact synonyms.
Can continuous learning reduce the cost of AI?
Potentially, but not automatically. Continual learning may reduce some of the data, computation and human intervention involved in periodically retraining models from scratch, and faster adaptation could reduce losses caused by deteriorating model performance. It also introduces additional technical and governance costs, including monitoring what the system learns, protecting existing capabilities and validating changes before they affect production decisions. Research continues to explore when the benefits outweigh those costs.
What is catastrophic forgetting?
Catastrophic forgetting occurs when an AI system learns new information or tasks and, in doing so, loses performance on knowledge or abilities it had previously acquired. It is one of the central technical challenges in continual learning because a useful adaptive system must learn from changing conditions without repeatedly destroying knowledge that remains valuable.
What is the difference between continuous learning and traditional machine learning?
Traditional machine-learning workflows commonly separate training from deployment: a model is trained on a dataset, evaluated and then released into production, with new versions created through later retraining. Continuous-learning approaches attempt to make learning more incremental, allowing new information or experience to influence the system after deployment. In practice, many real systems sit somewhere between these two extremes and combine monitoring, periodic retraining, retrieval, online updates and human supervision.
Are continuous-learning AI systems already being used?
Commercial development is underway, although approaches vary significantly. Wakeline describes a continuous-learning architecture for live decision systems, Trajectory is building infrastructure intended to improve AI models from real product interactions, and Lightwheel is developing continuous-learning loops for Physical AI and robotics. These systems should not be assumed to represent a single mature industry standard: continual learning remains an active technical field and many of the strongest performance claims currently come from the companies developing the technology themselves.
What should businesses measure instead of cost per token?
Cost per token remains useful for estimating model consumption, but it is a weak measure of overall economic value. For operational AI, businesses may learn more by looking at the cost of achieving a reliable outcome and, over longer periods, the total cost of maintaining reliable performance. That calculation can include model usage, human intervention, infrastructure, monitoring, retraining and the financial effects of incorrect decisions.