AI freedom and cost control: What happens when tokens spiral?
Anthropic and OpenAI are horribly unprofitable. Ed Zitron.
It may sound like a deliberately provocative headline, but the concerns surrounding these frontier model companies and the economics of artificial intelligence are no longer confined to the usual AI doomers on X. Ed Zitron, one of the industry's most persistent critics, has been arguing for some time that OpenAI sits at the centre of a much larger problem. The AI economy is being fuelled by extraordinary amounts of capital and compute, while the underlying economics of the business remain far from settled.
There is now another number worth considering. OpenAI has revealed that some of its researchers are consuming more than $7,000 worth of AI tokens every day, while median daily usage across its research organisation rose from $162 in July to more than $600 by mid-August.
OpenAI appears comfortable with this level of consumption because it believes the productivity gains generated by increasingly capable AI can justify the cost. Its researchers are using AI to write code, conduct research, run experiments, troubleshoot problems and perform increasingly sophisticated tasks. The objective is not simply to give people another productivity tool. It is to create systems capable of undertaking substantial amounts of work themselves.
That development has implications well beyond the economics of an AI company. It raises a question that agencies involved in digital production should be starting to consider: what happens when the cost of intelligence becomes a significant component of the cost of production?
For the web industry in particular, this could become an important distinction.
When AI becomes part of production economics
For the past couple of years, the conversation around AI has largely centred on productivity. Agencies have been encouraged to use AI to write code, generate content, conduct research, produce designs, test websites, automate repetitive tasks and accelerate development. The basic proposition has been relatively straightforward: if AI allows people to accomplish more in less time, the cost of producing digital work should fall.
Agentic AI changes that calculation because it is capable of doing considerably more than simply responding to a prompt.
An agent can be given an objective and then pursue that objective through a large number of individual actions. It can consume context, make decisions, call tools, inspect its own work, try alternative approaches and continue working until it considers the task complete. As these systems become more capable, the amount of computation required to accomplish a task can increase significantly.
The relationship between capability and consumption therefore becomes important. A more capable system can explore more possibilities and undertake more work, but each additional action can require additional tokens and computation. The very characteristics that make agentic AI attractive for production can also increase the cost of using it.
There is an additional consideration. As AI systems become better at reasoning about problems, they can identify alternative approaches and suggest additional pathways that might improve an outcome. That can be valuable, but it can also result in substantially more activity taking place behind what appears to the user to be a relatively simple request.
AI is therefore moving from being something that sits alongside production to becoming part of production itself. Once that happens, token consumption, model selection and the cost of accessing intelligence become operational considerations for an agency in much the same way as developer time, hosting and software licences already are.
This makes the question of who controls those economics considerably more important.
The web platform as the gateway to intelligence
Most mainstream visual web platforms are increasingly incorporating AI into their products. From an agency perspective, the attraction is obvious. Having AI available inside the same environment used to design, build and manage a website makes the technology easy to access and potentially easier to adopt.
The issue is not whether that convenience is useful. It is the relationship it creates between the agency and the platform.
When the AI is supplied by the web platform, the platform effectively becomes the gateway between the agency and the intelligence being used to produce the work. The agency is no longer simply paying for access to a development environment. It may also be consuming AI through that environment, using the models and commercial arrangements selected by the platform.
That distinction may not matter greatly when AI usage is relatively inexpensive. It becomes considerably more significant when agentic systems become responsible for substantial portions of web production and the volume of tokens being consumed increases.
The platform is then sitting between the agency and an increasingly important production input. If the platform also participates financially in the AI usage generated by its customers, there is an obvious commercial incentive to consider how that consumption is monetised.
This is why agencies should start thinking about AI economics as part of their technology strategy rather than simply accepting whatever AI functionality happens to be included in their web platform.
The question is not whether a platform has AI. The more important question is who controls the relationship with that AI.
A different approach to AI freedom
This is one of the reasons Blutui has taken a different approach.
The objective is not to create another closed environment in which an agency's access to AI is determined by the platform. Instead, Blutui is designed to give agencies control over how they connect intelligence to their production environment.
MCP technology is integrated throughout the platform, allowing front-end developers to connect to Blutui MCP and use agentic capabilities to build, develop and manage projects while retaining the ability to work with the AI model that best suits the task.
That could mean using a free model for a relatively straightforward piece of work, a specialist model for a particular technical problem or a more sophisticated reasoning model where the additional capability justifies the cost. Different models can be appropriate for different jobs, and the model landscape is changing quickly enough that an agency should not assume the best choice today will remain the best choice tomorrow.
This flexibility becomes increasingly valuable as the AI market evolves. Agencies can change models when better technology becomes available, take advantage of falling prices or adopt open models as their capabilities improve without having to rebuild their entire web production environment around a particular AI provider.
The principle is simple: the agency chooses the intelligence, the agency controls the tokens and the agency decides what level of AI consumption makes economic sense for each piece of work.
Keeping the AI economics transparent
There is another important distinction in Blutui's approach. Blutui does not meter, mark up or take a share of an agency's AI usage.
That means the platform does not have a financial incentive to increase the amount of AI consumption generated by an agency. The agency's platform costs remain predictable, while the cost of the intelligence it chooses to use remains under its own control.
This matters because AI economics are likely to continue changing. Model prices have already moved rapidly, while capability has increased and usage has expanded. A model that represents good value today may not do so in the future, and a model that is currently too expensive for routine production may become economically viable as prices fall.
Agencies therefore need the ability to respond to those changes rather than having their AI strategy determined by the commercial model of their web platform.
AI freedom is ultimately about preserving that choice. Cost control follows from having the ability to choose where and how intelligence is consumed.
The importance of retaining the code
There is a second element to this question that becomes particularly important as AI becomes more deeply embedded in web production: control over the underlying code.
Blutui does not place a barrier between an agency and the front end of the websites it produces. Front-end developers can work directly with the code whenever the project requires it.
That provides a useful degree of independence from the AI layer itself. AI can accelerate development, assist with implementation and automate significant amounts of work, but it does not become the only mechanism through which the agency can produce or modify its websites.
That distinction may become increasingly valuable if the economics of AI change significantly.
Token prices could increase for particular models. A provider could change its pricing structure. A model could become prohibitively expensive for certain workloads. A new generation of agentic development tools could radically alter the relationship between developer time and machine consumption.
An agency operating entirely inside an AI-powered visual production environment could find that its options are constrained by the platform through which that intelligence is delivered.
An agency that retains direct access to its front end has another option. Its developers can continue working directly with the code while using AI as an accelerator rather than becoming entirely dependent on it.
The technology can change without taking control of the production environment away from the people responsible for delivering the work.
Preparing for an uncertain AI economy
Nobody knows precisely where the economics of AI are heading. Models may become dramatically cheaper as competition increases and infrastructure becomes more efficient. Open models may become capable enough to challenge the commercial leaders. At the other end of the market, premium reasoning models may remain expensive because the computational resources required to operate them continue to increase.
Agentic development could also consume substantially more tokens than today's AI-assisted workflows because systems will be capable of undertaking much larger bodies of work autonomously.
All of these possibilities are plausible, and the speed at which the market is developing makes long-term predictions particularly unreliable.
For agencies, the sensible response is therefore not to try to predict which model or provider will ultimately win. It is to avoid unnecessary dependency on any particular outcome.
The technology infrastructure supporting web production should allow agencies to take advantage of whatever becomes available, whether that means a cheaper model, a more capable model, an open model or a completely different approach to delivering machine intelligence.
That requires control.
An agency should be able to choose the model that makes sense for a particular task, change that choice when the economics change and retain access to the underlying code regardless of which AI systems are being used.
The future of AI may be uncertain, but the need for control over the technology used to run an agency's business is not.
That is ultimately what Blutui means by createControl. It is not simply a statement about giving developers access to code. It is a broader principle about ensuring that agencies retain control over their production environment as the technology underneath it evolves.
The agency should decide which intelligence it uses. It should decide how much it is prepared to spend on that intelligence. It should be able to change its approach as the market changes, without having to abandon the platform on which its production operation is built.
And its developers should always be able to get to the code.
AI is going to change the economics of digital production. The agencies that benefit most may not necessarily be those that use the most AI, but those that retain the greatest freedom over how they use it.
Your models, your tokens, your choice.
Use the intelligence you want, control the cost and keep control of the code.
That is createControl.