Registry / No-training default

Which AI providers don't train on your API data?

As of 19 August 2026, of 16 AI model offerings tracked, 13 have clear public evidence, 1 is conditional or gated, and 2 have no public evidence.

Clear public evidence: OpenAI API, Azure OpenAI Service, Anthropic API, Claude via AWS Bedrock, Claude via Google Vertex AI, Gemini via Vertex AI, AWS Bedrock (platform), Mistral via Azure AI, Cohere via AWS Bedrock, Llama via AWS Bedrock, Llama via Azure AI, xAI API, DeepSeek via Fireworks AI
Conditional, sales-gated, or partial: Mistral La Plateforme
No public evidence: Cohere API, DeepSeek API (first-party)

Public commitments not to train on customer API data by default, cell by cell, with sources. The cell answers: Is there a public commitment not to train on customer API data by default? Statuses below are evidence grades, not endorsements, “no public evidence” means we could not verify it from public sources, not that the answer is no.

OpenAI API first-party API

●Yes, public confidence: high · verified 2026-08-15

OpenAI's own Help Center page publicly states that API data is not used for training by default, confirming the commitment.

source · full cell

Azure OpenAI Service OpenAI model, served by Microsoft Azure

●Yes, public confidence: high · verified 2026-08-14

The commitment is openly published on Microsoft's own Learn page with no gate, NDA, or portal login, so the default no-training commitment is publicly verifiable: yes_public, not yes_sales_gated. Verbatim from learn.microsoft.com: "are NOT used to train any generative AI foundation models without yo

source · full cell

Anthropic API first-party API

●Yes, public confidence: high · verified 2026-08-17

Verified verbatim on Anthropic's own privacy domain, covering the Anthropic API by name; a public commitment with no gate, so yes_public is correct.

source · full cell

Claude via AWS Bedrock Anthropic model, served by AWS Bedrock

●Yes, public confidence: high · verified 2026-08-18

The AWS Bedrock FAQ publicly states that customer content, including API data, is not used to train or improve the underlying models, covering Claude via Bedrock.

source · full cell

Claude via Google Vertex AI Anthropic model, served by Google Cloud Vertex AI

●Yes, public confidence: high · verified 2026-08-19

Google's publicly available data‑governance page explicitly states a default no‑training commitment that applies to all managed models on Vertex AI, including Claude.

source · full cell

Gemini via Vertex AI Google model, served by Google Cloud Vertex AI

●Yes, public confidence: high · verified 2026-08-18

The public Google Cloud docs state a training restriction covering all Gemini models, confirming a public commitment not to train on customer API data by default.

source · full cell

AWS Bedrock (platform) platform row

●Yes, public confidence: high · verified 2026-08-14

AWS Bedrock’s own privacy page publicly commits that customer prompts and outputs are not used to train the models by default, meeting the public commitment criterion.

source · full cell

Mistral La Plateforme first-party API

◔Partial confidence: high · verified 2026-08-14

Mistral's own help page confirms the quote verbatim and shows the no-training-by-default commitment is conditional on the pay-as-you-go plan, making "partial" the accurate fit.

source · full cell

Mistral via Azure AI Mistral AI model, served by Microsoft Azure

●Yes, public confidence: high · verified 2026-08-16

Microsoft’s public FAQ for Azure Foundry Models explicitly states that customer data is not used to retrain models, covering Mistral via Azure AI.

source · full cell

Cohere API Cohere model, served by Cohere (first-party)

○No public evidence confidence: high · verified 2026-08-14

Cohere's own enterprise-data-commitments page confirms an opt-out model — prompts/generations may be used for training unless the customer opts out ("If you are opted out, prompts and generations will not be used to train Cohere models") — so no public commitment not to train by default exists, matc

source · full cell

Cohere via AWS Bedrock Cohere model, served by AWS Bedrock

●Yes, public confidence: high · verified 2026-08-16

The AWS Bedrock FAQ publicly states that customer content is not used to improve base models nor shared with providers, covering Cohere models accessed via Bedrock.

source · full cell

Llama via AWS Bedrock Meta model, served by AWS Bedrock

●Yes, public confidence: high · verified 2026-08-19

The AWS Bedrock FAQ publicly states that customer content is not used to improve base models, confirming a default no‑training commitment for Llama on Bedrock.

source · full cell

Llama via Azure AI Meta model, served by Microsoft Azure (Azure AI Foundry / Models-as-a-Service)

●Yes, public confidence: high · verified 2026-08-19

The provider's own public documentation explicitly states that prompts and outputs are not used to train or improve any models, confirming a public no‑training commitment for Llama on Azure AI.

source · full cell

xAI API xAI model, served by xAI (first-party)

●Yes, public confidence: high · verified 2026-08-14

The provider's own public FAQ page explicitly states that xAI does not train on API inputs or outputs by default, confirming a public, ungated commitment.

source · full cell

DeepSeek API (first-party) first-party API

○No public evidence confidence: high · verified 2026-08-16

The privacy policy states DeepSeek uses data to train models by default and only offers an opt‑out, providing no public commitment not to train on customer API data.

source · full cell

DeepSeek via Fireworks AI DeepSeek model, served by Fireworks AI

●Yes, public confidence: high · verified 2026-08-19

The provider's own privacy notice publicly states they do not train on API inputs without explicit opt‑in, satisfying the commitment.

source · full cell