Every model price list has a second, invisible price list printed underneath it: the data terms. A free tier is not a cheaper version of the paid tier. It is a different product, paid for in a different currency — your inputs. I know, because I audited my own stack and found my cheapest channel was the one with the broadest rights to my data.
When you run a one-person company on an AI fleet, model selection is not a procurement detail. Your prompts are your operating system: drafts, decisions, negotiations, the raw material of the business, flowing through whichever API you wired up last. Choosing that API by price per million tokens is like choosing a bank by the color of its app. The number is real. It is just not the specification that decides what happens to you.
Here is what I used to get wrong: I treated an API key as a privacy boundary. If it came through a developer console instead of a consumer chat window, I assumed the data was handled like a business contract. It is not the key that decides. It is the billing tier behind the key. The same provider can run two completely different data regimes under two tiers of the same API — and the free one is not the generous one.
"We don't train on your data" answers only half the question
Data terms have two independent axes, and vendors hope you will conflate them. The first is training: will your inputs and outputs be used to improve the provider's models? The second is retention: how long does the provider keep them, for abuse monitoring or anything else, and who can look at them while they sit there?
A provider can honestly say "we never train on API data" and still retain your content for weeks of monitoring. Another can promise short retention and still reserve training rights somewhere in the consumer terms. These are separate promises, negotiated separately, with separate exceptions. When I review a channel now, I ask two questions, in this order: Do you train on this, by default, on this tier? And: How long do you keep it, and is there a zero-retention path — and which models are excluded from it?
The free tiers fail this review in writing. Google's unpaid-services terms for its Gemini API say, in plain language, that content submitted to the unpaid services is used to improve and develop Google's products and machine-learning technologies, that human reviewers may read, annotate, and process API inputs and outputs, and — set apart as a warning — that you should not submit sensitive, confidential, or personal information to the unpaid services. That is not fine print sleight of hand. That is the product description. The $0 is not a discount. It is an exchange rate: data for inference.
The real cost of a call is price plus surrender, weighted by sensitivity
Do the arithmetic honestly and the ranking inverts. For a throwaway call — summarize a public article, draft a joke — the free tier's data terms cost you almost nothing, because the input was worth almost nothing. For the call that carries your actual business, the same terms are the most expensive option on the list, no matter what the paid tier charges. Real cost equals sticker price plus data surrender multiplied by the sensitivity of what you sent. At $0 sticker price and high sensitivity, "free" is the priciest line item you own.
This is why the cleanest tier in my current stack surprised me. The channel I verified as my independent middle layer — Cloudflare's Workers AI — publishes the strongest default I found: no training on customer content, and zero retention of inputs and outputs for standard inference, free tier included, presented not as an enterprise upsell but as the built-in default. Meanwhile several famous "cheap" options turned out, on a terms read, to default the other way. The price table and the terms table ranked the same vendors in opposite orders.
Two more traps complete the picture. First, terms follow the pipe, not the endpoint: route a call through a gateway or aggregator and the gateway's terms govern your data, whatever the final provider promises. Second, terms are a function of time. Providers revise them unilaterally; a tier you audited in spring can be a different product by winter. A terms audit is a quarterly chore, not a launch checklist item.
So my selection procedure is now backwards from everyone else's. Terms first, on two axes, tier by tier. Price second. Benchmarks last — because a model that keeps what it should not keep is not a bargain at any speed. If you run anything real through model APIs, read your cheapest channel's terms tonight. The most dangerous line in your stack is probably the one that costs nothing.
— Jiahui Miao, founder of vertciti, building a personal AI operating system for one.
Top comments (1)
I think the problem is that the industry default is 'help us make it better', as opposed to 'keep it private, opt in to help'. But dont think it's just LLM providers, look at windows, per-install, you need to opt out, then they still force you to use a microsoft account, which guess what, needs to be opted out too of collecting user data, then there's your browser, where they nerf the experience, unless you allow them to collect data (eg. Gemini, YT), when local storage is a thing, private cloud is a thing, they easily COULD allow you to keep all your data secure and private, but they choose not to, because that's how they power their ad-tech. LLMs are just the more 'personal' version of it all, yet if you think about it, what does Google do with your gmail emails? Or Meta with your whatsapp data? "E2E encrypted', yet it takes a 4 digit pin to unlock your archive? Most of the world's information systems are built for data harvesting, because 'free' just means you cant quantify the cost yourself, it doesnt make it non-existent. That's kinda why open-source and self-hosting has kicked off so hard now, because it's physically impossible to get any privacy otherwise.