Free AI Tools That Are Actually Free (And What They Cost)

Genuinely free AI tools exist, and three of them do not train on what you type: Groq, Ollama running locally, and Gemini Notebook. Everything else on this list either trains on your inputs by default or asks you to choose. Twelve free tiers were checked against each vendor's own pricing page, terms, and privacy documents in August 2026. The tools that could not be verified are named at the end, not guessed at.

Key takeaways

• Seven of the twelve free tiers state in their own terms that inputs are used to train or improve models; Claude makes it your choice.
• Opt-outs are forward-only. Anthropic and ElevenLabs both state the switch does not reach data already collected.
• Google keeps human-reviewed Gemini chats up to 3 years, disconnected from your account, and does not delete them when you delete your activity.
• Saying yes to Anthropic model improvement extends retention from 30 days to five years — that is a priced trade, not a checkbox.
• Groq, local Ollama and Gemini Notebook are the three picks where the free tier does not buy your data.

What free actually costs#

Every roundup tells you the limit. The limit is the advertised price — 100 prompts, 2,000 completions, 10k credits. It is the number the vendor is comfortable putting on a marketing page.

The real price is in a different document — the clause in the terms or privacy policy that says what happens to the text, code and audio you send. For seven of the twelve tiers below, that clause says the vendor uses your inputs to build its next model. An eighth, Claude, puts the decision to you.

That is not a scandal. It is a trade, and a reasonable one for a haiku or a regex. It is a bad one for a client contract, an unreleased product spec, or a customer's support ticket.

So read the table with three questions in mind. What is the actual limit? What happens when you hit it? And what does the vendor do with what you typed?

The twelve free tiers, side by side#

Free AI tiers verified against vendor pricing pages, terms and privacy docs, August 2026

Tool	The actual free limit	What happens at the limit	Trains on your inputs?
Claude Free	No numeric cap published; 200k context; Fable gets 50% of the weekly limits other models get	Not published	Your choice. Allow it and retention goes to 5 years; decline and it stays 30 days
Gemini app (free)	No numeric prompt cap published; Gemini 3.6 Flash plus limited Gemini 3.1 Pro, Deep Research, Gems, 15 GB storage	Not published	Yes. Turn off Keep Activity and future chats are excluded unless you send feedback
Gemini API free tier	Free tokens on a long model list including Flash, Flash-Lite, Gemini 2.5 Pro, embeddings and Gemma 4; rate limits are shown in AI Studio, not published as fixed numbers	Spend-based caps return 429 RESOURCE_EXHAUSTED	Yes. The pricing page free-tier row reads content used to improve products
Gemini Notebook	Not stated on the data policy page	Not stated	No, unless you submit feedback on an interaction
GitHub Copilot Free	2,000 completions and 50 chat requests per month	Not published	Yes, unless you opt out in GitHub account settings
Grammarly Free	100 AI prompts per month, against 2,000 per member per month on Pro	Not published	Yes. On by default for Free accounts; you can opt out yourself
Groq	Per model, and not published as fixed numbers. The docs show Developer plan limits and send you to the limits page in your account settings for your own	Higher limits require the Developer plan	No. The Services Agreement bars it, with no free-tier carve-out
Hugging Face (free account)	$0.10 in monthly inference credits, subject to change; Spaces CPU Basic at 2 vCPU and 16 GB; ZeroGPU listed as free	You can buy additional credits and keep going	Not addressed on the pages checked
ElevenLabs Free	10k credits per month, roughly 10 minutes of speech at 128 kbps	Starter at $6/month raises it to 30k credits	Yes, with an opt-out in the Data use menu — and it is forward-only
Mistral Le Chat (free)	Unpublished. Limited messages, web searches and coding sessions, subject to fair usage	Not published	Yes, subject to an opt-out you set in your account
DeepSeek	Not published on the verified page	Not published	Yes, with a stated right to opt out; data is stored in the PRC
Ollama (local)	Unlimited, no account required, on macOS, Linux and Windows	There is no cap — you are limited by your own hardware	No. Prompt and response data is never logged or trained on

Two patterns fall out of that table. First, the vendors most willing to publish a hard number are the ones with the least generous tier: GitHub tells you exactly what 2,000 completions means, while Google and Anthropic publish no number at all for their far more useful consumer chat apps.

Second, the tools that do not train on your inputs are not the famous ones. They are an inference provider, a local runtime, and a research notebook.

Note what the "actual free limit" column does not contain: a single hard number from any of the three privacy picks. The vendors that protect your inputs are not the ones that tell you how much you get.

The three free tiers that do not train on your inputs#

These are the only three in this set where the vendor's own documents say your content stays out of the training pipeline.

The free tiers that train unless you say no#

Seven tiers state that they use your inputs, and Claude asks you to decide. Six of them give you a switch, and every switch lives somewhere different. The seventh, the Gemini API free tier, has no toggle at all — there the only control is the paid tier, which the next section covers.

1. Grammarly Free — account.grammarly.com/security/privacy

   Grammarly's support documentation states that Product Improvement and Training is on by default for Free, Premium and single-user Pro accounts. Free and Premium users can turn it off themselves at the page above. Multi-user Pro accounts bought through the website are also on by default, but only administrators can change it.

2. GitHub Copilot Free — GitHub account settings

   GitHub states it may use interactions from Copilot Free, Pro and Pro+ users — inputs, outputs, code snippets and associated context — to train and improve its models unless the user has opted out. The opt-out lives in your GitHub account settings and can be changed at any time. If you write code for a client under an NDA, do this before you install the extension.

3. Gemini app — turn off Keep Activity

   Google uses Gemini Apps data to provide, develop and personalize its services, and states those uses extend to the generative AI models. With Keep Activity off, future chats are not used for training unless you send feedback, and temporary chats are kept 72 hours. With it on, chats auto-delete after 18 months by default, adjustable to 3 or 36 months, or never.

4. Claude Free — the model improvement setting

   Anthropic asks consumer users to choose whether their chats and coding sessions train new models, with a retention consequence attached that the next section covers. The policy applies to Claude Free, Pro and Max only — not to Claude for Work, API access, or its Government and Education products.

5. ElevenLabs — the Data use menu

   ElevenLabs' privacy policy says it may process personal data to research, develop and train its AI models, including audio, text, video, preferences and feedback. The opt-out sits in the Data use menu under Terms and Privacy in your account. The policy draws no distinction between free and paid plans here.

6. Mistral Le Chat — the account-level objection control

   Mistral's privacy policy says it uses input and output data to train its models, subject to your opt-out, and that it has introduced a user control letting you object to that use directly from your account. The policy does not differentiate free from paid tiers.

DeepSeek belongs in this group too. Its privacy policy states inputs are used to improve and develop the services and to train its models, and grants users a right to opt out of that use.

Three things the opt-out toggle does not do#

Flipping the switch is necessary and nowhere near sufficient.

It is forward-only#

Anthropic's help page on changing the model improvement setting is unusually direct: turning it off leaves your data in model training that has already started, and in models already trained. Anthropic stops using previously stored chats in future runs, which is real protection — but nothing already in a run comes back out.

ElevenLabs says the same in policy language. Its opt-out "will only apply with respect to Personal Data provided" after you submit it. Two vendors, two documents, one rule: the toggle governs your future, not your past.

It does not reach human-reviewed copies#

Google states that a subset of Gemini Apps chats is read by human reviewers, including trained service providers, with the chats disconnected from your Google account first. Those reviewed copies are retained separately for up to three years — and Google states plainly they are not deleted when you delete your activity.

Your delete button does not reach the reviewed copy. The same pattern appears in the Gemini API terms, where human reviewers may read and annotate input and output on unpaid services after it is disconnected from your account and API key.

It does not cover safety-flagged conversations#

Anthropic's privacy documentation lists three routes by which consumer data reaches training: you allow it, you opt into a specific program, or your conversation is flagged for safety review. The third does not ask permission. Incognito chats are excluded even when the setting is on, which is the cleanest control Anthropic offers.

Free tiers that are really trials in disguise#

A free tier refreshes on a schedule and has no end date. A trial is a one-time grant with a clock on it. Four of the best-known "free" offers in AI are the second thing.

One-time grants that are commonly presented as free tiers

Offer	What you actually get	How it ends
Runway free plan	125 one-time credits that do not refresh monthly, plus 5 GB asset storage. Gen-4.5 costs 60 credits per 5 seconds of video	The credits do not expire, but at 60 credits per clip that is roughly two videos, ever
Google Cloud Free Trial	$300 in Welcome credit to spend over 90 days, with no automatic billing	Miss the 90 days or exhaust the credit and the billing account closes, stopping all associated projects and resources
AWS Free Tier credits	Up to $200 in credits — $100 up front, $100 more earned by exploring services	The account closes on its own six months after you open it, or when the credits run out
Grammarly Pro trial	Seven days of Pro, which carries 2,000 prompts per member per month against the Free plan 100	The trial runs 7 days; Grammarly does not publish the post-trial behavior on its plans page

Runway is the clearest case. The credits genuinely never expire, which sounds generous until you divide 125 by the 60 credits a five-second Gen-4.5 clip costs. That is an evaluation budget, not a free tier.

Google and AWS both attach a shutdown to their headline number. Google Cloud gives a 30-day grace period to recover resources if you upgrade within the window, the only safety net of the four. AWS is blunter: the account closes on its own at six months.

Why the paid API is often more private than the free chat#

For several vendors, paying moves you into a stricter data regime, not just a bigger one.

OpenAI's API documentation states that data sent to the API is "not used to train or improve OpenAI models" unless you explicitly opt in. Abuse monitoring logs are kept up to 30 days by default, audio outputs for an hour, and prompt caching data up to 24 hours. That page addresses the API throughout and makes no statement about consumer ChatGPT plans.

Google draws the same line and puts it in a pricing table. The free Gemini API tier row states content is used to improve products; the paid tier row states it is not. On paid services Google does not use prompts — including, in its own list, system instructions, cached content and files such as images, videos or documents — or responses to improve its products.

Anthropic draws it too. Its consumer terms update explicitly excludes Claude for Work, API access, and its Government and Education products from the consumer training choice.

The pattern is consistent enough to act on. If you are handling anything you would not post publicly, the developer tier of the same product is usually the cheaper privacy purchase. The guide to estimating LLM API costs covers what that switch runs per month.

Where you are changes the answer#

Google's Gemini API terms state that in the European Economic Area, Switzerland and the United Kingdom, free services are treated as paid services for data handling. Same free tier, same API key, two different privacy outcomes depending on where you sit. In Berlin your free-tier prompts get the paid-tier treatment; in Boston they do not.

DeepSeek's privacy policy states that it directly collects, processes and stores personal data in the People's Republic of China, and that data tied to an account is kept as long as the account exists. That is a jurisdiction question, not a quality question, and it decides the tool for some organizations before any evaluation begins.

The honest limitation on every pick#

Three of those deserve a sentence more.

Hugging Face gives free accounts $0.10 in monthly inference credits, marked subject to change, against $2.00 for PRO at $9/month (checked August 2026). The free Spaces hardware is the better deal: CPU Basic at 2 vCPU and 16 GB, plus ZeroGPU listed as free. As of July 2025 the hf-inference provider focuses mostly on CPU work such as embeddings, ranking and classification.

ElevenLabs Free gives 10k credits a month, about ten minutes of speech at 128 kbps and 44.1 kHz. The commercial license is the detail that catches people: it begins at Starter, from $6/month (checked August 2026). If you want that audio in a client video, the free plan is not the plan. The transcription roundup covers the reverse direction.

Claude Free is the most capable no-cost chat tier in this set — 200k context, web search, memory across conversations, file creation and code execution, Slack and Google Workspace integrations, and access to Fable, Opus, Sonnet and Haiku. Anthropic publishes no usage number beyond noting that Fable receives 50% of the weekly limits other models get. Pro is $17/month billed annually, or $20/month monthly (checked August 2026).

Who this is not for#

If you are shipping a product on a free tier, stop reading and go price the paid one. Google explicitly reserves the right to modify its free tier with 30 days' notice. A dependency you cannot pay to keep is not a dependency.

If you work in a regulated function — health records, legal files, anything with a data processing agreement behind it — free consumer tiers are the wrong shelf entirely. The clauses above are written for individuals, and none of these consumer terms offer the processor commitments a compliance team will ask for.

And if you need guaranteed throughput, none of these will give it to you. Free tiers are best-effort by design. Both Groq and Google send you to your own account page for the authoritative figures rather than publishing free-tier numbers, which tells you how firm those numbers are meant to be.

What could not be verified, and dropped#

ChatGPT is not in the table. No OpenAI page carrying a ChatGPT free-tier limit or a consumer training default is cited in the sources above, so no such claim is made here. The single OpenAI claim in this article is the API one, and it comes from the API data page.

Microsoft Copilot, Canva, Perplexity and Adobe Firefly are absent for the same reason: no verified vendor page, so no row. LM Studio states its cloud services are "Zero Data Retention (ZDR) across the board" and that voice and audio data "is processed locally and never leaves your device," but publishes no pricing or commercial-use terms on that page, so it stays in prose rather than the table.

Several vendors decline to publish the figures that would make planning possible. Google will not put fixed RPM, TPM or RPD numbers on its Gemini API rate limits page, directing you to AI Studio instead. Groq's rate limits page publishes Developer plan numbers and sends free users to their own account settings. Neither Google nor Anthropic publishes a numeric cap for its consumer chat free tier, and Mistral uses only relative language. GitHub, Grammarly and Google all state a limit without stating what happens when you reach it.

Before you paste anything in#

Five checks, once per tool, before the first prompt that matters.

1. Find the training clause, not the FAQ

   Search the privacy policy for "train" and "improve." Marketing pages say what the vendor wants; the policy says what it may do.

2. Set the opt-out before your first real prompt

   It is forward-only at every vendor in this list. Do it on account creation, not on your first sensitive task.

3. Check whether the control is yours to set

   On multi-user Grammarly accounts bought through the website, only administrators can change the training setting. Assume the same structure elsewhere until you confirm otherwise.

4. Read the retention number attached to your choice

   Anthropic prices its consumer choice at 30 days versus five years. Google keeps human-reviewed Gemini chats up to three years regardless of your deletions.

5. Ask whether it is a tier or a grant

   If the allowance does not refresh on a stated schedule, it is a trial. Runway, Google Cloud credits and AWS credits all fail that test.

Run the full vendor review from the AI tool security checklist for anything that will touch customer data, and use the pre-purchase evaluation guide before a free tier turns into a subscription line item.

The verdict#

Use Ollama locally for anything you would not email to a stranger. Ollama states that running models on your own hardware is always unlimited, that no account is needed to download it, and that prompt and response data is never logged or trained on. The cost is your own hardware.

Use Groq for free API inference. Section 4.2 of its Services Agreement bars training on inputs and outputs with no free-tier exception, which is a stronger commitment than any other hosted tool here makes. Check your own limits page before you design around it — Groq does not publish free-plan numbers.

Use Claude Free or the Gemini app for general chat, and treat both as public. Neither publishes a usage number. The Gemini app uses your chats for model improvement until you turn Keep Activity off; Claude puts the same decision to you as a choice. Set the control on day one, accept that it does not reach what you already sent, and keep sensitive work somewhere else.

Choosing a coding assistant specifically? The coding assistant comparison goes deeper on Copilot Free and its alternatives. And if a free tier is standing in for a workflow you will eventually pay for, price it now — the paid developer tier is frequently the cheaper privacy decision, not just the faster one.