Is subscription a suitable business model for an AI product? Is AI getting more expensive? Are tokens getting more expensive? Or are we just moving toward a fair model?

If you make people pay based on how they use a service, it is well known that they then use it little and cannot realize its full value - they never really learn it, because they keep counting money and feeling afraid. Subscription solved this brilliantly in the past - it pulls you into the game, you use it, get used to it, do not want to lose it. And you spread it around, increasing the user base. In short, flat subscription is ideal for adoption.

With ordinary chat, of course, a very active user consumes far more tokens and therefore generates costs not 10x between the average and power user, but certainly more like 100x. Maybe the business model can still handle that, but some restrictions may be used:

• Limits within an hour
• Temporary switch to a cheaper model
• Slower response (not deliberately, but perhaps because someone overshooting is deprioritized)

But that is chat - agents change everything. So-called Deep Research was already a bit of a problem, but a full-fledged agent is much bigger - and in coding it is the only thing that makes sense today. An agent can work on a given task for hours, load code and documentation (so it has enormous token context), call tools across hundreds of interactions, create colleague subagents and let them work on parts of the problem in parallel. And if your agent works on something for an hour, will you watch it, or open another window and start another one solving something else? Of course you will do that, so in the end a de facto unlimited number of hungry agents and subagents is running. The ratio between your average user and Power User can easily reach a 4-order-of-magnitude difference and can practically go completely off the rails - not intentionally at all, the user simply used agents fully and sleep is not a natural limit.

A solution typically has to include payment for the fact that it exists - someone built it, handles its security, architecture, compliance, and also has some basic operating costs. But that is not what we care about now - we want to focus on two sources of cost that are directly measurable and directly depend on the user (without the user they do not arise).

I think this is the model of the future - a subscription that unlocks features and typically also includes some prepaid consumption, ideally pooled. Let us explain.

GitHub Copilot (announced and in effect from June) or Cursor work on the principle that you have a subscription that lets you use its features and at the same time includes some amount of consumption. That amount may even be absolutely zero (seat-only, as Anthropic Claude Enterprise has it, with nothing included in the price), or smaller (for a 20 USD license you get 10 USD included), but the standard is usually from 1:1 upward (Cursor has 1:1 on the 20 USD license, but the 60 USD plan includes 70 USD of consumption; GitHub Copilot uses 1:1, so a 19 USD license includes 19 USD of consumption).

What I find very fair, and as far as I know only GitHub Copilot and Cursor have it, is the pool. All purchased licenses create a pool of money for the whole company, so a user who uses it little does not mean their consumption is lost - someone else uses it. Ideal for AI infusing into an organization - everyone should have it! This is a good incentive: nobody is "not promising" and therefore without AI; everyone has it, and if someone does not use it fully, someone else consumes it. Of course it is important to have good ways to manage possible per-user limits and top-up policy by company or its organizations - GitHub Copilot has this handled very well.

Claude Code is very readable in the Enterprise version - the license includes absolutely nothing, you buy everything extra. I think the era of vague limits (week, hour, we will never say exactly, or we will say it but then keep changing it and granting exceptions) will end, especially for enterprise, because that kind of fog is not good for companies. OpenAI Codex, however, is still currently operating this way.

Currently all of this concerns coding agents, but other AI services such as M365 Copilot, Gemini, OpenAI and Claude still typically run on a more classic subscription with strange limits. Claude Enterprise is an exception - even for ordinary chatting, nothing is included in the price; it is seat-only. I guess this is a preview that one day other enterprise services will also move into a different model, and I expect something quite similar to coding agents - a fee per hour of borrowed computer (for example for Cowork), and a fee for tokens.

What about you? Do you prefer a flat subscription world with strange and unclear limits, subtle quality reduction (see some Claude affairs from the last few days), but potentially better price predictability and the ability to get several orders of magnitude more out of it if you are a power user? Or do you prefer a model based on actual consumption with a separate entry-ticket fee? Personally, because I grew up on Azure and consider pay-as-you-go the most natural and fair model, I like the model GitHub Copilot has now - subscription costs something, but you get its value back in tokens and it is pooled, so you do not have to be afraid to give a subscription even to people who do not yet know how to use it that much.

So is AI getting more expensive? I do not think so - tokens across the whole market are priced very fairly with a reasonable margin and never get more expensive for a specific generation/quality. Rather, new models arrive, and in practice today's mini model in the cloud gives you a similar service at 5x lower price than a high-end model a year ago; the fee per unit of intelligence is therefore dropping absolutely dramatically, more than in any other IT technology over the last 50 years. But - today a coding agent writes a whole application for you in an hour; a year ago an AI assistant suggested a new class, and when you ran it, you dealt with the fact that it did not work and sent it back. If living in the past is enough for you, use a 5x cheaper model and enjoy saving money - AI has become brutally cheaper. I would rather ride a different wave - for very similar money to a year or two ago, today I have a substantially better and more useful solution. I prefer intelligence and growth over savings. But everyone can have it differently, and that is fine - but AI has not become more expensive. However, we have hit the limit of a business model that is unsustainable in this era. In my view we need a model where, for predictable costs, we introduce AI to everyone, do not lose unused tokens thanks to pooling, and sensibly manage top-up limits, for example by organization or individually.

I have prepared and tested tips for saving tokens - we will look at that in the next article!