Which AI APIs Train on Your Data? Free Tiers, Paid APIs and Retention, Checked Against 18 Providers' Own Terms (October 2026)

Last updated: October 18, 2026

Short answer: free tiers are where training happens, paid APIs mostly say they do not train on your data, and retention is a separate question that "no training" does not answer. Google uses unpaid Gemini API traffic to improve its products unless you are in the EEA, Switzerland or the UK. Mistral's Free mode "may use your data" for training until you flip an opt-out toggle. OpenRouter's free models train or not depending on which provider serves them. On paid tiers, 14 of the 18 providers we checked say no training by default, three (Mistral, Cohere and DeepSeek) do not state the default, and one (Moonshot) says both yes and no on two of its own pages. Separately, most keep prompts for around 30 days for abuse monitoring, and a handful keep nothing by default.

Everything below comes from two directories we keep on this site, each built from the providers' own terms, docs and privacy pages with verbatim quotes: the free API directory (17 providers) and the data retention directory (18 providers). Every claim links to the provider's page it came from. Where the two directories differ for one provider, it is usually because they cover different tiers or read different documents, and we say which.

The table#

The 18 rows are the providers in the retention directory. The "free tier" column comes from the free API directory; "Not covered" means that directory does not list the provider. Provider names link to our page for that provider, which carries the full quotes.

One note on the Groq and Cerebras rows. The retention directory marks their training as "not stated" because the pages it read (Groq's Your Data page, Cerebras's privacy policy) do not address training. The free API directory read different documents that do. Groq's Services Agreement says "Groq is not permitted to use Inputs or Outputs for training or fine-tuning any AI Model Services or other models, unless explicitly granted permission or instructed by Customer." Cerebras's Terms of Service say they do not grant "Cerebras the right to use Service Content for the purpose of training or fine-tuning models." Both documents cover the whole service, not just the free tier, so the table shows "No" with the source named.

The free tier is where training happens#

Google. The Gemini API terms are the clearest case of a free tier paid for with data: "When you use Unpaid Services, including, for example, Google AI Studio and the unpaid quota on Gemini API, Google uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services and machine learning technologies." The same terms add that human reviewers "may read, annotate, and process your API input and output" and say plainly: "Do not submit sensitive, confidential, or personal information to the Unpaid Services."

The exception is geographic: "If you're in the European Economic Area, Switzerland, or the United Kingdom, the terms under 'How Google uses Your Data' in 'Paid Services' apply to all Services, including Google AI Studio and unpaid quota in the Gemini API, even though they are offered free of charge." Switch to the paid quota and the terms flip: "Google 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." The full record is on our Google free tier page.

Mistral. Mistral's help center says: "Free mode: As stated during subscription, we may use your data (input and output) to train our artificial intelligence models. You have the right to opt out of this program at any time." The opt-out instructions point to Admin panel, Privacy, "Anonymous improvement data", and warn that "Vibe and API opt-out toggles are separate." One more line applies on every plan: thumbs up or down plus a comment "authorize us to use your rating, your input and output in order to improve the behavior of our artificial intelligence models." See our Mistral free tier page.

OpenRouter. OpenRouter itself "does not store your prompts or responses" unless you opt in, but whether a free model trains depends on who serves it. Its logging docs say: "Wherever possible, OpenRouter works with providers to ensure that prompts will not be trained on, but there are exceptions. If you opt out of training in your account settings, OpenRouter will not route to providers that train." Of the 12 free models in our snapshot, six are served by providers OpenRouter lists as "May train" (Nvidia and Thinking Machines). The training setting is split: "There are separate settings for paid and free models." One caveat from our OpenRouter free tier page: OpenRouter's provider table lists Google AI Studio as "Does not train", which is OpenRouter's characterisation and differs from Google's own unpaid-quota terms quoted above.

The rest of the free tiers say no, or say nothing. Groq and Cerebras say no in the documents quoted above. Cloudflare Workers AI says "Cloudflare does not use Customer Content processed by Workers AI to train AI models." Alibaba Cloud says it "will not use your Member Content to develop or improve the models on Model Studio, unless you separately provide your consent." Z.ai's terms say no for "enterprises and developers using API Services" unless you explicitly agree, but the same terms reserve the right to use content from "individual users" to improve services. Cohere is the gap: its enterprise data commitments apply to "customers who have a credit card on file", while "Our Terms of Use and Privacy Policy apply to data we receive from users using a trial API key." It does offer a dashboard opt-out. NVIDIA's trial terms do not mention training but say NVIDIA "will not store or use User Content or Generated Content at the end of each API Service session", with exceptions.

OpenAI has no free text tier, and its nearest thing is paid in data. OpenAI's help center says "Some organizations may qualify for daily complimentary tokens on traffic shared with OpenAI", and shared data "will be used to help identify usage patterns, measure model quality, and inform future evaluation and training of models." You also need "a positive account balance to use OpenAI models and the free tokens."

If you are picking a free tier for a hobby project, our access guides for DeepSeek V4.1 Flash and GLM-5.3 list the routes. Read this section next to them before you paste anything you would not post publicly.

Paid APIs mostly say no, with exceptions#

On the paid side the default is no training. OpenAI's data controls guide: "As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)." Anthropic's Commercial Terms: "Anthropic may not train models on Customer Content from Services." xAI's security FAQ: "SpaceXSI never trains on your API inputs or outputs without your explicit permission." The clouds say the same: Bedrock ("AWS and the third-party model providers will not use any inputs to or outputs from Amazon Bedrock to train Amazon Nova, Amazon Titan, or any third-party models"), Vertex AI ("Google won't use your data to train or fine-tune any AI/ML models without your prior permission or instruction") and Azure (prompts and completions "are NOT used to train any generative AI foundation models without your permission or instruction"). Together AI and Fireworks AI say no in their own docs and terms too.

The exceptions are where the wording stops short:

• Mistral (paid). The commercial terms say no training "except (a) when you (i) opted-in to training on a Mistral AI Product set to opt-out by default or (ii) have not opted-out of training on a Mistral AI Product set to opt-in by default", plus feedback, Order Forms and "Labs or Preview Models". Which way the API toggle is set by default for pay-as-you-go is not stated. Check the toggle yourself.
• Cohere. "You can opt out from your prompts and generations being used to train Cohere models in your dashboard settings at any time." The default state is not stated.
• DeepSeek. The Open Platform terms do not address training. The privacy policy lists "to train and improve our technology, such as our machine learning models and algorithms" as a purpose for inputs, and also says it does not cover data from apps developers build on the open platform. That leaves the API question open.
• Moonshot. Its terms say content "may be used" unless you have a written agreement. Its help center says the opposite. Details below.
• Together AI does not train by default, but its privacy docs say "By default, Together stores the prompts you send and the responses models return, and may use them for product improvements." Training and product improvement are different words, and both are on the page.

Retention is a separate question#

"We don't train on it" does not mean "we don't keep it." Most providers keep prompts and responses for a fixed window to look for abuse:

• About 30 days: OpenAI ("abuse monitoring logs are generated for all API feature usage and retained for up to 30 days"), Mistral ("thirty (30) rolling days to monitor abuse"), xAI ("stored on our servers (encrypted at rest) for 30 days for auditing purposes"), Alibaba Cloud ("no longer than 30 days") and Cohere ("We automatically delete logged prompts and generations after 30 days").
• Longer: the paid Gemini API keeps data "for fifty-five (55) days" per its usage policies. DeepSeek and Moonshot tie retention to the account: DeepSeek keeps personal data, including input, "for as long as you have an account."
• Only when flagged: Vertex AI logs prompts only "if automated safety classifiers detect suspicious activity", for up to 90 days, and customers with a Google Cloud Master Agreement "are exempt from prompt logging for this abuse monitoring by default." Azure may store flagged prompts and completions for human review; the pages read do not give a duration.
• Not retained by default: Groq ("By default, Groq does not retain customer data for inference requests"), Cerebras ("We do not retain inputs and outputs associated with our training, inference and chatbot Services"), Fireworks ("Fireworks has Zero Data Retention by default" for open models), Z.ai (content "is not saved on our servers") and OpenRouter, which still stores request metadata such as token counts and latency.

Zero data retention (ZDR) is how you close the 30-day window, and access varies a lot. It is self-serve on Groq ("All customers may enable Zero Data Retention (ZDR) in Data Controls settings"), Bedrock (data_retention_mode none, per Region), xAI (team admins, "where available"), Together ("ZDR is not enabled by default") and OpenRouter as a routing filter. It needs approval at OpenAI, Anthropic, Vertex AI and Mistral, and an enterprise relationship at Azure, Alibaba Cloud and Cohere. The Gemini API ZDR page sends you elsewhere: "If your workload requires guaranteed zero data retention or enterprise data processing agreements, use Vertex AI."

ZDR has carve-outs everywhere. Anthropic's retention docs say Fable 5, Fable 5.1, Mythos 5 and Mythos 5.1 "require 30-day data retention and are not available under ZDR unless expressly authorized by Anthropic", and flagged chats may be kept "for up to 2 years." We covered what that did to existing agreements in Fable 5 broke enterprise ZDR agreements, and the Bedrock side in Fable 5 on AWS Bedrock.

Where the docs disagree#

Four providers publish two pages that say different things. We show both sides and do not pick one.

Anthropic, on default retention. The privacy center: "For Anthropic API users, we automatically delete inputs and outputs on our backend within 30 days of receipt or generation." The API data retention docs: "Conversation content (your prompts and Claude's outputs) is not retained by default; the exception is Covered Models, which require 30-day retention." Both are Anthropic's own pages. Neither says Anthropic trains on the data.

Groq, on default retention. The body of the Your Data page says "By default, Groq does not retain customer data for inference requests." The same page's meta description still says "standard 30-day retention", which reads as stale against the body. The body also says reliability and abuse logs, when they exist, "are retained for up to 30 days."

Mistral, on who can get ZDR. The docs: "ZDR is available on paid plans for supported stateless API calls." The help center: "Zero Data Retention (ZDR) is available only with pay-as-you-go and only for stateless API calls." The help center adds: "We will review your request and, at our discretion, approve or deny it."

Moonshot, on training. The OpenPlatform terms: "Customer who requires restrictions on the use of Customer Content for training or improving Moonshot AI models may contact Moonshot AI to discuss available enterprise arrangements or separate written agreements. Unless otherwise expressly agreed in writing, Customer Content may be used for the foregoing purposes." The Kimi help center: "No. User data submitted via the API, including input content and model outputs, will not be used to train or improve Kimi models." The help page is served for the China locale and its examples use api.moonshot.cn, so it may describe the China platform rather than the overseas one. Until Moonshot reconciles them, the terms are the safer reading.

What to do#

1. Pay for anything sensitive, or get ZDR. The single biggest change in the table is free versus paid. On Google it is the whole difference. If you need nothing kept at all, pick a provider where ZDR is self-serve or the default (Groq, Bedrock, Together, OpenRouter's ZDR routing, Cerebras, Fireworks for open models) or ask for it where it needs approval.
2. Turn the toggles off yourself. Mistral (Admin panel, Privacy, "Anonymous improvement data", and the separate Vibe toggle), Cohere (dashboard Data Controls) and OpenRouter (separate free and paid training settings) all have opt-outs whose defaults are either "on" or not stated. On Mistral, rating a response with the thumbs up or down button and adding a comment authorizes use of that exchange on any plan.
3. Check the region endpoint, not the company's headquarters. Default endpoints usually have no fixed location: xAI's global endpoint "may route requests between regions", and on Vertex AI "Requests submitted to a global endpoint may be processed in any Google Cloud location around the world." Regional options exist but cost more or need approval: Anthropic's inference_geo: "us" "is priced at 1.1x the standard rate" on Claude 4.6 and later; OpenAI charges "a 10% uplift" on eligible newer models, and non-US regions require approval for abuse monitoring controls; Mistral offers EU and US regional endpoints, though "Stateful features, including Agents, Batch, and the Files API, are not available on regional endpoints"; OpenRouter's eu.openrouter.ai and us.openrouter.ai are on the Business and Enterprise plans. Some providers state one location: DeepSeek stores data "in People's Republic of China", Moonshot "in Singapore", and Z.ai says customer data "is generally processed in Singapore."
4. Turn off stateful features you do not use. Retention windows quoted above are for plain inference. Stored state runs longer: OpenAI's Responses API keeps data "for at least 30 days" by default or with store=true, and /v1/conversations keeps it until deleted; Gemini's Interactions API stores interactions by default (55 days paid, 1 day free) and its Files API keeps files for 48 hours; Fireworks keeps Responses API conversations for 30 days with store=True, its default; Anthropic's batch API keeps data up to 29 days; Mistral's Agents API keeps input and output "until you terminate your account." Set store=false, delete files, and skip conversation storage unless you need it.
5. If you move a workload, check compatibility first. Swapping to a provider with better terms is easier when it accepts the same request shape; our API compatibility reference tracks that.
6. Or keep the data on your machine. Running open-weight models locally means your prompts never leave your hardware, so none of the above applies. Our local models directory covers what fits on which hardware, and the license guide covers what each model's license lets you do. For coding specifically, see the best local coding LLMs.

This is a reading of published terms, not legal advice. Terms change, enterprise contracts override them, and your obligations depend on your own data and jurisdiction.

FAQ#

Does the Gemini API free tier train on my data?#

Yes, outside the EEA, Switzerland and the UK. Google's terms say unpaid quota content is used "to provide, improve, and develop Google products and services and machine learning technologies", and human reviewers may read it. Users in those three regions get the paid-service terms even on unpaid quota.

Does the OpenAI API train on my data?#

Not by default. Data sent to the API since March 1, 2023 "is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)." Opting in is the condition for OpenAI's daily complimentary tokens.

Does Anthropic's Claude API train on my data?#

No. The Commercial Terms say "Anthropic may not train models on Customer Content from Services." Retention is a separate question, and Anthropic's own two pages describe it differently.

Which AI APIs keep no data by default?#

On the pages we read: Groq, Cerebras, Fireworks AI (for open models), Z.ai and OpenRouter (which still keeps request metadata). Groq's page metadata still mentions 30-day retention, which reads as stale.

Is "no training" the same as zero data retention?#

No. OpenAI, Mistral, xAI, Alibaba Cloud and Cohere all say they do not train on paid API data by default (or offer an opt-out) and still keep prompts around 30 days for abuse monitoring. ZDR is what removes that window, and it usually has exclusions for stateful features.

Continue Reading#

• Free AI API Directory - free tiers, limits and data use for 17 providers, each with the quote behind it.
• AI API Data Retention Directory - retention, training, ZDR and residency for 18 providers, with the "docs disagree" notes.
• Fable 5 Broke Enterprise ZDR Agreements - the Covered Models exception in practice.
• Where to Run DeepSeek V4.1 Flash Free and Cheap - free and cheap routes, to read alongside the training terms above.
• The Best Local Coding LLMs in 2026 - the option where no provider sees your prompts at all.

Sources#

All read on October 18, 2026, via the two directories above. Provider pages quoted:

• Google: Gemini API terms, usage policies, ZDR
• Vertex AI: abuse monitoring, zero data retention, data residency
• OpenAI: your data, data sharing help article
• Anthropic: Commercial Terms, privacy center retention article, API and data retention, data residency
• Azure: data privacy
• Amazon Bedrock: FAQs, data retention
• Mistral: training help article, opt-out help article, commercial terms, privacy policy, ZDR docs, ZDR help article, regional inference
• xAI: security FAQ, regions
• DeepSeek: privacy policy
• Moonshot AI: OpenPlatform model use terms, user privacy, Kimi API data security
• Z.ai: terms of use, privacy policy
• Alibaba Cloud: product terms, Model Studio data retention policy
• Cohere: enterprise data commitments
• Groq: Services Agreement, your data
• Cerebras: Terms of Service, privacy policy
• Together AI: privacy and security, zero data retention
• Fireworks AI: data handling
• OpenRouter: logging, data collection, ZDR, in-region routing
• Cloudflare: Workers AI data usage
• NVIDIA: API trial terms

Verified October 18, 2026.