What does 'free AI' actually cost you?

"How to use AI for free" is a tutorial question, and a good one, answered elsewhere in this network in practical step-by-step terms. This piece asks a different question: when a price reads £0, what is actually happening economically, and what does "free" quietly leave out?

The honest answer is that almost nothing in AI is free in the sense that a public library book is free. Someone is paying for the electricity, the data centre, the model training run, or the attention and data you hand over in place of cash. Where that cost actually lands, not just whether a sign-up form asked for a card number, is the difference between a genuinely useful choice and one you regret once you are dependent on it.

Three different things get called "free"

Lumping every free AI option together hides more than it reveals. In practice there are at least three distinct categories, and they carry very different economics.

• Open-weight models you run yourself. Meta's Llama family, Mistral's open releases, Google DeepMind's Gemma models, Microsoft's Phi models, Alibaba's Qwen models and DeepSeek's V4 models are all published with downloadable weights that carry no licence fee for most uses. Downloading them costs nothing. Running them well is a separate question.
• Free tiers of hosted commercial products. ChatGPT, Claude.ai, Gemini and Microsoft Copilot all offer a version you can use without paying, typically with a capped number of requests on their strongest model, a default step down to a smaller or older model once that cap is hit, and narrower access to features such as file uploads or extended context than the paid tier gets.
• Free apps funded by something other than a subscription. Meta AI, built into WhatsApp, Instagram and Facebook, charges the user nothing because Meta's revenue comes from the advertising and engagement economics of those platforms, not from a subscription to the assistant itself.

Only the first category is free in a fairly straightforward sense, and even there, "free" refers to the licence, not to the electricity, hardware or time it takes to actually use the model.

The compute cost behind "open"

A small open-weight model, the kind released in a few-billion-parameter size specifically so it can run on ordinary hardware, will genuinely run on a decent laptop's processor or a modest consumer graphics card, at no cost beyond electricity you were already paying for. That is a real, usable form of free.

The larger, more capable open-weight releases are a different proposition. DeepSeek's V4-Pro model and Meta's largest Llama 4 model are both mixture-of-experts designs, DeepSeek's running to more than a trillion parameters in total and Meta's to several hundred billion, even though only a fraction of either activates for any single request. Running a model at that scale at a usable speed generally means either several high-end graphics cards bought outright, a serious capital cost, or renting equivalent hardware from a cloud provider by the hour, a running cost that scales with use. Free-to-download and free-to-run are not the same claim, and the gap between them is exactly where the real economics of "open" AI live.

There is a middle path. Hosted services such as Hugging Face's free inference tools or Groq's free API tier let you send requests to an open-weight model without owning any hardware, usually within a rate limit generous enough for personal experimentation and tight enough to make clear you are not meant to run a business on it.

The data and privacy cost

The free tiers of hosted commercial products are not funded purely out of goodwill. It is common, though not universal or fixed, for a free consumer tier's terms of service to reserve the right to use your conversations to help improve the underlying model, while the same company's paid API or enterprise tier usually carries a stronger commitment not to train on customer data by default. Exactly which providers do what, and under what opt-out settings, changes often enough that the only reliable check is the current privacy terms of the specific product you are about to use, not a general rule repeated in an article.

This matters economically, not only ethically. If a free tool's business model depends on your usage data, you are paying with information rather than money, and the question worth asking is whether what you are handing over, client details, unpublished work, anything you would not want to see reflected back in someone else's output, is worth more to you than the fee you are avoiding.

The time cost and the lock-in cost

Two further costs rarely appear in any comparison. The first is time: a rate-limited free tier that quietly drops you to a weaker model once you hit its cap will often need more prompting, more correction and more patience to get a usable answer, and that extra effort is a real cost even though no invoice ever records it. The second is lock-in: the prompts you have refined, the chat history you have built up and the habits you have learned in one product's interface do not transfer cleanly to another, so the longer you stay on one free tool, the more it costs, in time and disruption, to leave it for a better or cheaper one later. Our piece on digital sovereignty's economics covers this switching-cost pattern in more general terms, and it applies just as much to a person choosing a free chatbot as it does to an organisation choosing a cloud vendor.

How to actually weigh the decision

None of this is an argument against using free AI tools. For most people, the first and often the right move is to explore with something that costs no money while working out whether a task genuinely needs AI at all. It is an argument for asking the right questions before treating "free" as the end of the analysis. How often will you actually use this, occasionally out of curiosity, or daily as part of real work? Does the work involve anything you would not want used to improve someone else's model? Do you already own hardware capable of running a smaller open-weight model locally, making that a genuinely near-zero-cost option? And if the free tier's limits eventually push you onto a paid plan, is that the plan you would have chosen deliberately, or just the one you drifted into?

A free tool that answers a genuine curiosity, at no meaningful cost in time, data or lock-in, is exactly the kind of access this publication's founder set out to protect: the ability to learn and build before a course fee, or any other fee, decides the question for you.

Disclosure

Common Fortune is a Mickai publication. Trust Agent, under the same ownership, offers free courses on using AI tools at trust-agent.ai. Naming it here is a disclosure, not a claim of neutrality. Mickai also sells AI-readiness advice to organisations at mickai.co.uk/ai-readiness, and currently offers free access to its own product through an invitation-only closed beta, free for the first 100 testers who apply, at mickai.co.uk/beta. Both are real options among many, and the economics set out above apply whichever tool, free or paid, you end up choosing.