Free AI tools compared: what each free tier really costs you
Free AI tools are free in the only sense that matters to your card: no money leaves the account. The cost lands somewhere else, and the vendors print it on their own pages. We read the pricing sheets, the rate limit docs and the data terms for six of them in August 2026, and the pattern held across all six. Where the free tier is metered, the vendor has mostly stopped printing the meter. Where it isn’t metered, the vendor is usually taking the content instead.
That’s the whole trade, which makes “which free AI tool is best” the wrong question. The better one is which free tier bills you in a currency you can afford. Below is what each vendor’s own pages said, the two places where free is genuinely free, and the one free tier we’d name as the loser.
The same three questions, asked of every vendor’s own pages
We didn’t run a capability bake-off, and we’re saying so up front because most free-tool roundups pretend that they did. The task here was documentary. We opened each vendor’s pricing page, its rate limit documentation and its data terms, then asked three fixed questions of every one: what does the free tier cost in money, what does it cost in throughput, and what does the vendor do with whatever you type into it.
Three questions, six products, same sources every time. The answers to the first are boring and identical, because $0 is $0. But the answers to the third are where these products stop resembling each other, so that’s where most of this piece sits.
| Free tier | Printed price | The limit the vendor prints |
|---|---|---|
| Claude Free | $0 | “Usage limits apply”, no number given |
| Gemini API free tier | $0 input and output | No table in the docs, view them in AI Studio |
| GitHub Copilot Free | $0 | 2,000 code completions a month, limited chat |
| Hugging Face free account | No price printed for it | Credits not printed; PRO gets 20x at $9/mo |
| Ollama Free | $0 | One model at a time in the cloud, plus models on your hardware |
| OpenAI API | None on the page we read | gpt-5-nano at $0.05 in, $0.40 out per million tokens |
One gap in that table is ours rather than theirs. OpenAI’s consumer pages returned an HTTP 403 to us on every attempt, so ChatGPT‘s free tier isn’t described here. We’d rather leave a hole than fill it from memory, because a term you half-remember is worse than a term you admit you couldn’t read.
Google’s free API tier is paid for in content
Google is the clearest of the six, and that clarity is the reason it comes off worst. The Gemini API pricing page lists a free tier against each of the five Flash models in the section we read, with free input and free output tokens. Beside every one of those free rows sits the same short note: “Content used to improve our products.”
The API terms spell out what that note is doing. For the unpaid services, Google says it “uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services”, and that “human reviewers may read, annotate, and process your API input and output”. Google says it disconnects that data from your account first. Then it gives the instruction that settles the matter.
Do not submit sensitive, confidential, or personal information to the Unpaid Services.
Gemini API Additional Terms of Service, Google
The same terms describe the paid tier completely differently. There, Google says it “doesn’t use your prompts (including associated system instructions, cached content, and files such as images, videos, or documents) or responses to improve our products”, and logs prompts only for a limited period to catch policy violations. Same model, same API, and the money is what buys the confidentiality.
Google’s consumer app carries a softer version of the same deal. Its privacy notice says a subset of chats are read by human reviewers, that reviewed chats are kept for up to three years even after you delete them from your activity, and that chats auto-delete after 18 months by default. Turn Keep Activity off and future chats aren’t used to train the models, though they’re still held for 72 hours. The notice asks you not to enter confidential information you wouldn’t want a reviewer to see.
Same model, same API. The money is what buys the confidentiality.
Rundowns AI
Anthropic charges for the free tier in years, not tokens
Anthropic prices the same trade on a different axis, and it’s the more interesting one because the default runs the other way. Its privacy centre says the company uses your chats and coding sessions to improve its models if you choose to allow it, and that Incognito chats are excluded even when you’ve enabled model improvement. So the content isn’t taken by default. What the choice changes is how long the company keeps what you sent.
The consumer terms update published on 28 August 2025 put numbers on it. The change covers the Claude Free, Pro and Max plans, including Claude Code used from those accounts, and existing users had until 8 October 2025 to choose. Allow training and new or resumed chats are retained for five years. Decline and you stay on the existing 30-day retention. Anthropic’s privacy article adds that a conversation you submit feedback on is stored for up to five years regardless.
Five years against 30 days is a factor of roughly 60 in exposure, by our arithmetic, and it’s the same toggle on the free plan as on the $100 one. That’s the honest shape of the offer: Anthropic isn’t charging free users in content, it’s charging everyone who opts in with a retention window. Which is a better deal than Google’s unpaid tier, and still a real one.
| Product | Is your content used to improve models? | Retention the vendor states |
|---|---|---|
| Gemini API, unpaid tier | Yes, and human reviewers may read it | Not stated in the terms we read |
| Gemini API, paid tier | No, per the same terms | Limited period, for policy enforcement |
| Gemini consumer app | A subset is reviewed by humans | 18 months default, 72 hours with activity off, 3 years if reviewed |
| Claude Free, Pro and Max | Only if you allow it | 5 years if you allow it, 30 days if you decline |
| OpenAI API | “We do not train our models on inputs and outputs through our API” | Not stated on the page we read |
| Ollama, run locally | Runs on your own hardware | Yours |
The rate limits stopped being published
Data terms are at least written down, so you can read them before you commit. Throughput mostly isn’t any more, though, and that’s the change worth noticing if you last checked a free tier a year or two ago.
Google’s rate limits documentation no longer carries a free tier table. It says limits are tied to a project’s usage tier and points you at AI Studio to view your own, so the number you’d plan a side project around is visible only after you’ve built the project. The doc does print the tier ladder: the free tier needs an active project, paid Tier 2 opens at $100 of cumulative spend and Tier 3 at $1,000, and moving off free requires setting up billing first.
Anthropic’s pricing page does the same thing in fewer words. The Free plan lists real capability, including web search, memory across conversations, file creation and code execution, extended thinking, and connectors through remote MCP. The limit is three words: “Usage limits apply.” A support article explains that usage depends on the length and complexity of your conversations, the features you use, the model you pick and the effort level you’ve selected. All true, and none of it a number.
Against that, GitHub Copilot looks unusually straight. Its free plan prints $0 and 2,000 code completions a month with limited chat and agent usage, and GitHub’s docs add that Copilot Free is for individual developers without organisation access and runs on auto model selection only. You know before you start what you get. That’s rarer than it should be, and it’s most of why Copilot Free earns a place in the stack for people who just want completions.
Where free is actually free, and what it costs instead
Two of the six aren’t really selling a free tier at all, which is why they come out best on the third question. Running open-weight models on your own machine takes the vendor out of the loop entirely, so there’s nothing to opt out of because there’s nothing being sent anywhere.
Ollama describes itself as a way to build with open models on your computer and in the cloud, and its pricing page puts the free plan at $0 with unlimited public models, CLI, API and desktop apps, and the line that matters here: run models on your hardware, keep your data private. Its cloud free tier is lighter, running one model at a time, with Pro at $20 a month and Team at $25 per seat with a five-seat minimum. Hugging Face works similarly. Its pricing page says many core Hub features stay free, and puts PRO at $9 a month for 20 times the included inference credits.
The bill still exists, though, it just moves onto your desk. You’re paying in GPU memory, electricity and the hours it takes to get a model quantised and serving properly, and the model that fits is smaller than the one an API would hand you. We worked through what that trade looks like in practice in our guides on running an LLM on your own machine and on hosted APIs versus self-hosting. Neither one concludes that local is cheaper. They conclude it’s cheaper at some volumes and always more private, which is a different claim.
What the paid tier actually buys
Because the free tiers are priced in terms rather than money, the paid sheets are the only place to see what the terms are worth. OpenAI’s pricing puts gpt-5.6-sol at $5 per million input tokens and $30 per million output, gpt-5.6-luna at $0.20 and $1.20, and gpt-5-nano at $0.05 and $0.40. Google prices Gemini 3.7 Flash at $0.75 input and $3.75 output per million through 31 December 2026, rising to $1.50 and $7.50 on 1 January 2027, with Gemini 3.5 Flash-Lite at $0.30 and $2.50.
Put those against the free tier and the arithmetic gets uncomfortable in a useful way. A million input tokens on Flash-Lite costs 30 cents, and paying anything at all moves you onto the terms where Google says it doesn’t use your prompts. So the confidentiality is cheap at hobbyist volumes, and plenty of people are handing over content rather than buy it. Our model pricing comparison goes through the spread between tiers properly.
The subscription side reads the same way. Claude Pro is $17 a month billed annually or $20 monthly, with Max tiers starting at $100, while Copilot runs $10 for Pro with $15 of monthly credits, $39 for Pro+ with $70, and $100 for Max with $200. Entry pricing has converged on $10 to $20 across the board, which is the point we made when we looked at which AI tools earn a place in a developer stack.
The loser, and the four things we couldn’t check
Naming one is the part a comparison owes you. The free tier we’d avoid is the Gemini API’s unpaid tier, and not because Google hid anything. It’s the opposite. Google wrote down that content goes into product improvement, that human reviewers may read your inputs and outputs, and that you shouldn’t send anything sensitive, so anyone using it for work with real data is ignoring a warning printed on the tin.
The one we’d keep is Copilot Free, for the narrow reason that it prints its limit, so you can plan around it. Ollama and Hugging Face win the data question outright, though they win it by handing you the hardware bill instead. Claude Free sits between them, because it pairs strong capability and an opt-in rather than an opt-out on training with no published limit at all.
Four gaps, stated rather than estimated. OpenAI’s consumer pages blocked us, so ChatGPT’s free tier terms aren’t in this piece. Google no longer publishes free tier request limits in its docs. Hugging Face doesn’t print the credit allowance a free account gets, only that PRO is 20 times more. And neither Anthropic nor GitHub prints a message or request number for its free plan, so the throughput you’d actually get is unknowable before you sign up.
What would change this piece is simple enough to watch for. A vendor republishing free tier rate limits as a table, Google extending its no-training term to the unpaid tier, or Anthropic moving its consumer default from opt-in to opt-out. Until one of those happens, the free tiers worth using are the ones that told you the price in something other than money.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
