AI's flat rate won't last forever, and maybe it shouldn't

For months we’ve lived through a rather strange situation: for 20, 100 or 200 dollars a month we can access Artificial Intelligence models that, measured against API prices, could end up consuming the equivalent of hundreds or thousands of dollars if used intensively. To a regular user it might look like a bargain. For those of us who come from the infrastructure world, it also looks like an anomaly that sooner or later will have to be sorted out.

An independent analysis circulating among power users and developers has put numbers to that feeling. Its authors bought several Anthropic and OpenAI plans, ran long programming tasks until they hit the weekly limits, and compared that consumption with what it would have cost using API prices. The conclusion is striking: a Claude Max 20x subscription at 200 dollars a month could allow usage equivalent to about 8,000 dollars a month in API. On ChatGPT Pro 20x, also at 200 dollars, the approximate equivalent could reach 14,000 dollars.

It’s worth clarifying from the outset that this does not mean Anthropic or OpenAI internally pay those 8,000 or 14,000 dollars for each intensive user. API prices include margin, infrastructure, availability, product, support and commercial costs. But the comparison helps us understand something important: the AI flat rate works because most users don’t push the service to the limit. If many of us start using agents for hours, working with entire repositories, requesting long analyses or automating development tasks, the economics change.

The real cost is hidden in the subscription

Subscriptions have always worked on a similar idea. A gym doesn’t expect every member to train two hours a day. A streaming platform doesn’t assume that every user will watch content non-stop. Nor does an AI provider calculate its margin assuming that each person will exhaust their usage limits every week.

The difference is that a query to an advanced model isn’t already-produced content distributed at low cost. Each response consumes compute. Every long context, every code iteration, every tool executed and every agent working for minutes or hours has a real cost in GPUs, memory, network, storage, energy and operations.

As of publication, I’ve used as a rough reference the European Central Bank exchange rate of June 12, 2026, with 1 euro equivalent to 1.1567 dollars. That puts 1 dollar at about 0.8645 euros. The figures below are rounded, do not include taxes, fees or billing differences by country, and should be read as an approximation.

The table is striking because it breaks the perception of “I pay a fee and that’s it.” In reality, we’re using an advanced compute service that today is packaged as a subscription to make it simpler, more mainstream and more attractive. The problem is that consumption isn’t linear. A user who asks the odd question is nothing like a developer who uses code agents all day long.

The key is average utilization. If most people use little, the commercial model holds up. If more and more users start squeezing their subscriptions for long tasks, the flat rate becomes a cross-subsidy: light users compensate for heavy ones.

API, subscription or your own models: three different ways to pay

The comparison with the API helps to see where the cost is. OpenAI publishes prices for GPT-5.5 of 5 dollars per million input tokens and 30 dollars per million output tokens in short context, while GPT-5.5 Pro rises to 30 dollars per million input tokens and 180 dollars per million output tokens. Anthropic, for its part, lists Claude Fable 5 at 10 dollars per million input tokens and 50 dollars per million output, and Claude Opus 4.8 at 5 and 25 dollars respectively.

Converted to euros, the order of magnitude is clearer:

When you look at these prices you understand why the labs increasingly separate end user, API, professional plans and enterprise. The subscription works very well for mass adoption. The API works for product, integration and consumption control. The enterprise model lets you negotiate terms, support, security and limits. They are different businesses, even though they all use the same raw material: inference.

For an individual user, paying 20 or 200 dollars a month can be dirt cheap if AI saves hours of real work. For a company that puts AI into internal processes, the question changes: how much does each task cost? Which model is used? What data leaves? What latency is acceptable? What happens if prices go up or limits change?

The uncomfortable part is that we may have to get used to paying more to use advanced AI. Not because the companies are evil, but because serving high-end artificial intelligence costs money. If a model helps us program, analyze contracts, review documentation, generate reports or automate complex tasks, perhaps paying 100, 200 or 500 euros a month isn’t outrageous if the return is clear.

What doesn’t seem sustainable is thinking we’ll always have near-unlimited access to the best models in the world for a low flat fee. It can happen for a while due to competition, growth strategy and falling costs. But if professional use takes off, someone will have to foot the bill.

The alternative: open models, your own infrastructure and more control

The other possibility is that not everything has to go through paying more to OpenAI, Anthropic, Google or xAI. We’ll increasingly have more open or open-weight models available that we can run on our own servers, in private cloud or in public cloud. Maybe they’ll be slower. Maybe they won’t always match the quality of the best proprietary model of the moment. But for many uses they’ll be enough.

This is a part of the debate that interests me especially. Not everything needs the most powerful model. Many enterprise tasks are repetitive, internal, bounded and measurable: classifying documents, extracting data, summarizing incidents, answering against a knowledge base, reviewing logs, preparing drafts, generating simple SQL queries, analyzing tickets or helping technical teams with internal documentation.

For those cases, a well-deployed open model can be far more interesting than an external API. Not only because of cost. Also because of privacy, compliance, operational control and sovereignty. If the data doesn’t leave your environment, you reduce dependence and can better adapt the system to your needs. That said: the cost doesn’t disappear. It just changes form.

Running AI on your own infrastructure means paying for servers, GPUs, storage, electricity, cooling, network, administration, monitoring, updates, security and technical time. In the public cloud something similar happens, although the cost turns into on-demand consumption. In private cloud or bare metal you can gain predictability and control, but you need to size it well.

The interesting part will be choosing wisely. For a critical task of complex reasoning it may be worth using the best commercial model available. For an internal classification or summarization task a smaller open model may be enough. For code, it may make sense to combine tools: a subscription for personal productivity, the API for metered flows and your own models for repeatable tasks.

The AI architecture to come will be hybrid. Just as many companies combine public cloud, private cloud, SaaS and on-premise systems, they’ll also combine proprietary models, APIs, open models, local inference and specialized services. The question won’t be “which AI do I use,” but which AI do I use for each task, at what cost, under what control and with what dependence.

The AI bill will force us to mature

During the first phase of generative AI, many of us have used these tools as if the cost of intelligence were almost invisible. We open a window, ask, iterate, test and carry on. That’s normal: the product is designed so we don’t think about tokens or GPUs. But companies that build real processes on AI will have to look at the bill more carefully.

That doesn’t have to be a bad thing. In the cloud world we already went through something similar. First came the fascination with elasticity. Then came the unexpected bills. Later came FinOps, observability, reservations, optimization, hybrid architectures and a more mature cost culture. With AI something similar will happen.

We’ll have to measure cost per task, not just cost per user. We’ll have to decide which model deserves each flow. We’ll have to cache context, avoid unnecessarily giant prompts, use small models when they suffice, limit agents, log consumption and compare results. Efficiency will once again be a technical virtue, not an accountant’s obsession.

We’ll also have to accept that good Artificial Intelligence can cost money. We’ve grown used too quickly to an artificial abundance. If a tool saves us real work, reduces errors or lets us create products that weren’t viable before, paying more can make sense. The important thing is not to get trapped in a blind dependence on flat rates that may change tomorrow.

That’s why my conclusion isn’t “everything will be more expensive” nor “we must flee the big platforms.” My conclusion is more practical: it’s worth preparing for a world in which AI will have several prices, several qualities and several deployment forms. Sometimes we’ll pay more for the best model. Sometimes we’ll use slower but sufficient open models. Sometimes the API will interest us. Sometimes we’ll prefer our own infrastructure.

The flat rate has been magnificent for discovering AI. For production, strategy and technological sovereignty, we’ll need something more serious: clear costs, alternative models, data control and the ability to decide where each part of our artificial intelligence runs.

Frequently asked questions

Are AI subscriptions subsidized?

In a way, yes, for the most intensive users. The model works because many users pay and don’t exhaust all the limits. If you compare maximum consumption with API prices, some subscriptions can allow equivalent usage far above the monthly price.

Does this mean OpenAI or Anthropic are going to raise prices?

Not necessarily, but it would be reasonable to expect more segmentation. The most powerful models, long agents, large context windows or professional features could become increasingly tied to higher-tier plans, credits, API or enterprise contracts.

Does it make sense to run open source models on your own servers?

Yes, in many cases. It can be interesting for sensitive data, predictable costs, regulatory compliance or repeatable internal tasks. But it’s not free: it requires hardware, operations, security, maintenance and technical capacity.

Which strategy seems most reasonable for a company?

Use a combination of options. Commercial models for complex tasks, the API when integration and measurement are needed, open models for repeatable internal uses, and an architecture that lets you switch provider or model without redoing everything.