Two stories from this week that touch on a question you should ask of any AI service: who processes your data, and where? Anthropic now offers Claude in a second way via Google Cloud, and Mistral has released new speech models that you can run in the EU or on your own servers. There is also a small privacy feature in ChatGPT with clear limits.
Claude on Google Cloud, now run by Anthropic
On 22 October, Anthropic made Claude Models on Google Cloud (Anthropic managed) generally available. You sign in and pay through Google Cloud, and usage counts towards your existing Google Cloud agreements. But Anthropic runs the service itself and is the processor of your data. You immediately get all features of the Claude platform, including beta features, with Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5.
There is one important limitation: for now, the service only has US and Global endpoints. The existing variant, Claude Models on Google Cloud (Google managed), remains available. Anthropic itself recommends that variant if you want Google Cloud as your processor or have strict requirements about the region where your data stays.
What does this mean for you? You won't see any difference on the invoice, but under the GDPR it is a different contract with a different processor. If you use Claude with personal data, for example in software development with client data in your code, choose the variant deliberately and record which one it is. How Claude scores on privacy is set out at /claude/ and in the GDPR check for software development.
Voxtral 3: Mistral's speech models, also hosted in the EU
This week Mistral released Voxtral 3: models for transcription (Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3 and a realtime version), a model for speech (Voxtral Mini TTS 3) and an API for voice agents. Transcription works in more than 50 languages, including the 24 official EU languages. According to Mistral, transcription starts at $0.0035 per minute, speech costs $13 per million characters and the voice agent $0.04 to $0.08 per minute. You can use the models via Mistral's EU-hosted API or run them on your own infrastructure. TestingCatalog also reports US and Global endpoints.
What does this mean for you? Conversations with customers or patients almost always contain personal data. A transcription model that runs in the EU or on your own servers makes the GDPR assessment a lot simpler than a service outside the EU. This applies especially to customer service and healthcare. Mistral already supplies models to professional tools such as Cognigy and Leexi. Whether they already use Voxtral 3 is not yet stated on their websites. You can see Mistral's privacy profile at /mistral/ and in the GDPR check for customer service.
ChatGPT: locked chats are not encrypted chats
OpenAI is rolling out locked chats for Free, Go, Plus and Pro. You put a conversation behind a four-digit passcode or Face ID, and it disappears from your regular history and search results. The feature is not end-to-end encrypted and changes nothing about how long OpenAI keeps conversations or whether they are used for training. If you reset the passcode, all locked chats and files are permanently deleted.
What does this mean for you? The lock helps against people looking over your shoulder at your phone, not against processing by OpenAI. Client data therefore still doesn't belong there, see the GDPR check for ChatGPT in mental healthcare.
In brief
- Claude via Google Cloud can now be accessed in two ways, with two different processors. Only the Google-managed variant mentions regional storage.
- Mistral's Voxtral 3 offers transcription and speech with an EU-hosted API or on your own servers.
- A locked chat in ChatGPT is a screen lock, not extra privacy.
All models and their privacy scores side by side: /modellen/ and /veiligheid/.
