Mistral AI revised its Commercial Terms effective August 5, 2026, with the main operational changes being: the company expanded the definition of Customer Data to include 'access to data' and 'coding environment'; restructured its AI training opt-out framework to be clearer about which products use training by default and which require opt-in; consolidated references to 'Labs Models' and newly defined 'Preview Models' as a single category ('Labs or Preview Models') with the same data-training rules; and removed a carve-out from its liability cap that previously protected customers against breaches of confidentiality obligations, narrowing the exceptions to liability limitations.
Commercial customers: More types of material customers provide to Mistral can now be used for AI training, not just text prompts and uploaded training files.
Commercial customers using experimental models: If you have turned off training on your main Mistral products, that setting will not prevent Mistral from training on data you submit to experimental models.
Commercial customers: Mistral's cap on what it will pay customers for damages is now stricter and covers confidentiality breach claims that were previously outside the cap.
The updated terms redefine when Mistral AI will use customer data and outputs for AI model training. Previously, the terms distinguished between free tiers (where training was opt-out) and Labs Models (where training always occurred). The revised language consolidates experimental models as 'Labs or Preview Models' and explicitly states that training opt-out preferences applied to other products do not carry over to these experimental models. Additionally, the terms now specify that customer data includes 'access to data' and 'coding environment,' expanding what qualifies as trainable material. The liability cap was narrowed to remove the carve-out protecting confidentiality breach claims, meaning liability limits apply more broadly. You can review which Mistral products are set to opt-in versus opt-out by default and avoid Labs or Preview Models if you do not want your data used for training.
→ Review the opt-in and opt-out defaults for each Mistral product you use to understand which ones permit training by default.
→ If you do not want your data used for training, avoid submitting material to Labs or Preview Models, as training opt-out settings do not apply to them.
→ Audit your own privacy disclosures and customer commitments to confirm they accurately reflect Mistral's revised training rules.
This is the 2nd significant Ai Training Rights change Mistral AI has made since ConductAtlas began monitoring.
Across all monitored documents, Mistral AI has made 5 significant changes.
5 of Mistral AI's significant changes have been classified as negative for consumers.
Expanded definition of trainable data and restructured opt-in/opt-out framework; training on Labs or Preview Models now occurs by default regardless of settings on other products.
Removed carve-out protecting confidentiality breaches from liability cap, narrowing exceptions to the cap.
This change record describes what was added, removed, or modified in the document. Analysis reflects what the updated agreement states or permits. It does not constitute a legal determination about enforceability. Applicability may vary by jurisdiction. Methodology
This change materially modifies data training authority and liability exposure for commercial customers. The restructured opt-in/opt-out framework now clearly delineates which products default to training opt-in versus opt-out, and explicitly prevents zero-retention settings from applying …
Regulatory exposure, obligation change, escalation trigger, board-ready language, and recommended action for legal and compliance teams.
Unlock the full institutional analysis — InsightConductAtlas provides verified policy intelligence sourced directly from platform documents. All analysis is intended to support, not replace, legal and compliance review. Record CA-C-004262.
Mistral AI updated its Additional Product Terms effective August 5, 2026, making operational clarifications to its third-party integration and product …
Mistral AI's terms, effective August 5, 2026, removed API access from consumer use and restricted Mistral AI Studio to business …
Mistral AI's model card was updated to add Shieldstral 1.0 to the list of related models displayed alongside existing models …
Get the research letter
Companies change their terms quietly. We read every version and catch what actually changed. One email a week on the changes that matter and what they mean.
Get alerted when this policy changes again, including what changed and why it matters.