Brello vs Claude: where your questions go

At a glance

Brello 1.0 and Claude differ in where a question is processed and in what happens to a chat afterwards.

Where your question is processed

Brello 1.0 processes your question on the phone. Claude sends it to Anthropic, whose privacy policy says personal data is transferred to its servers in the US or to other countries.

In Brello 1.0, the prompt goes only to the on-device runtime, Google’s LiteRT-LM, which runs one of three downloaded models on the phone’s GPU, or on its CPU when the GPU isn’t available: Brello Pro, based on Gemma 4 E4B by Google; Brello Vision, based on Gemma 4 E2B by Google; or Brello Core, based on Qwen3 1.7B by Alibaba. All three are released under Apache 2.0, and Brello isn’t affiliated with or endorsed by Google or Alibaba. Photos are copied into Brello’s private app storage and never uploaded. There is no Brello server, proxy or relay. The one exception, web search, is covered below.

Anthropic’s privacy policy calls what you send “Inputs”: the content you submit, “including via third-party applications, services, and content you choose to upload, integrate or interact with using our Services”. Inputs generate responses and actions, which it calls “Outputs”. It says that when you use its services, “your personal data is transferred to our servers in the US, or to other countries outside the European Economic Area (‘EEA’) and the UK”, for example “to provide you with the Service, or to train our models, or to conduct research”.1

The policy covers Claude.ai and Anthropic’s other services, but “does not apply to content that we process on behalf of customers of our business offerings, such as our Enterprise accounts”.1 For the general difference between cloud and on-device assistants, see on-device AI, explained.

Model training and the model-improvement setting

Anthropic may use your Claude chats to train its models unless you opt out, and asks new users to choose during sign-up. Brello 1.0 has nothing to train on, because chats stay on the phone.

The privacy policy, effective 10 September 2026, says: “We may use your Inputs and Outputs to train and improve Anthropic AI models, unless you opt out through your account settings.” Opting out has two exceptions: Anthropic “will use Inputs and Outputs for model improvement” when conversations are “flagged for safety review”, or when “you’ve explicitly reported the materials to us (for example via our feedback mechanisms)”.1 Anthropic’s privacy centre adds a third route, opting in explicitly, for example by joining its Trusted Tester Program. Training data includes “the entire related conversation”, but not raw content from connectors such as Google Drive unless it is copied into the conversation.4

The setting dates from Anthropic’s update of 28 August 2025, which applied to Claude Free, Pro and Max. New users choose “in the signup process”, and existing users had to make a choice by 8 October 2025 to keep using Claude. Anthropic says it uses “a combination of tools and automated processes to filter or obfuscate sensitive data”.3 If you turn the setting off later, Anthropic “will not use your previous or new chats or coding sessions for future model training”, though your data “will still be included in model training runs that are already in progress, or in models that have been trained”.5 Incognito chats are never used, “even if you have enabled Model Improvement”.4

Brello 1.0 has no account system, no analytics and no crash reporting, and no Brello server receives your questions. We have no chats to train on.

How long conversations are kept

Claude keeps chats in your history until you delete them, then removes them from Anthropic’s back end within 30 days. With model improvement on, Anthropic may keep de-identified copies for up to 5 years.

Anthropic says a conversation you delete is “removed from your chat history immediately” and “deleted from our back-end storage systems within 30 days”. If you allow your chats to be used to improve Claude, “we may retain your data in a de-identified format for up to 5 years in our model training pipelines”, for new or resumed chats after the setting is turned on. Feedback you submit, for example with the thumbs up or down buttons, is kept for 5 years. In all cases Anthropic may keep chats “as required by law, to resolve disputes, or as necessary to combat violations of our Usage Policy”.5 If you don’t allow training, the 28 August 2025 update says you “continue with our existing 30-day data retention period”.3

Brello 1.0 writes chats, including any “Thought process” text, to a single file, conversations.json, in its private app storage. Android backup is turned off for the app, and its data-extraction rules exclude every storage domain from cloud backup and from device-to-device transfer. A chat stays until you swipe left to delete it, or delete every chat with “Delete all chats” in Settings. Nothing is synced, so if the phone is lost or reset, the chats are gone.

Flagged conversations

Anthropic uses conversations its systems flag for safety review even if you have opted out of training, and keeps their inputs and outputs for up to 2 years. Brello 1.0 has no server-side review of any kind.

Anthropic’s privacy centre says: “We retain inputs and outputs for up to 2 years and trust and safety classification scores for up to 7 years if your chat or session is flagged by our automated trust and safety systems as violating our Usage Policy.”5 Flagged conversations may be used “to improve our ability to detect and enforce our Usage Policy, including training models for use by our Safeguards team”.4 The privacy policy says flagged content is disassociated from your user ID to train Anthropic’s trust and safety models, but that Anthropic “may re-identify the Inputs or Outputs to enforce our Terms of Service or Usage Policy with the responsible user if necessary”.1

Brello 1.0 has no server, so nothing is flagged, kept or reviewed outside the phone. Its safeguards run in the app; for example, its system prompt tells the model to admit uncertainty. Safety at Brello describes them.

Accounts and sign-in

Brello 1.0 has no accounts. Anthropic’s terms say it may ask you to create one, and signed-in chats sync across your devices.

Anthropic’s consumer terms, effective 8 October 2025, say: “To access our Services, we may ask you to create an Account.” Users must be at least 18, or older where local law requires. If you sign up with an email address owned by your employer or another organisation, the account “may be linked to the organization’s Anthropic enterprise account”, and its administrator “may be able to monitor and control the Account, including having access to Materials”, which the terms define as your Inputs and Outputs. Anthropic says it gives notice before linking.2 Anthropic’s download page says: “Your conversations, projects, memory, and preferences sync across all devices when you’re signed in.”8

Brello 1.0 asks for nothing: there is no sign-up, and the models download from Hugging Face without an account or token.

Offline use

Brello 1.0 works offline after a one-time model download. None of the Anthropic pages we checked describes using Claude without a connection.

Brello Core is a 977 MB download, Brello Vision 2.59 GB and Brello Pro 3.66 GB, and Wi-Fi is recommended for it. After that, the assistant works fully offline; only web search, which is optional, needs a connection.

Web search: who sees the query

Claude can search the web, and in Anthropic’s new Claude experience it decides when to. Brello 1.0 searches only when you allow it, directly from the phone.

Anthropic’s help centre says you can have Claude “search the internet to provide you with up-to-date information”, and that “if you have the new Claude experience, there’s no web search toggle. Claude searches the web when it helps.” With web search on, Claude can also fetch the full content of a page whose URL you give it. The article says image search “is powered by Bing”, but doesn’t say which provider handles text searches or what is sent to it.7

Brello 1.0’s web search is off by default. With it off, a question that looks time-sensitive brings up a card titled “Search the web for this?” with two buttons, “Search the web” and “Answer offline”. When Brello searches, the phone sends a cleaned-up search text directly to DuckDuckGo, Bing, Brave Search, Google News or Wikipedia, then fetches up to four result pages itself, with Do Not Track and Global Privacy Control headers. The model that writes the answer still runs on the phone. The search engine and the websites see a normal request, including the phone’s IP address. How AI assistants handle your data explains these patterns in general terms.

What Claude offers that Brello 1.0 doesn’t

Claude does far more than Brello 1.0. These are the differences Anthropic’s own pages describe.

• More platforms, kept in sync. Apps for the web, macOS, Windows, iOS and Android, with conversations, projects, memory and preferences synced when you’re signed in.8 Brello 1.0 runs on Android phones and iPhone and syncs nothing.
• Memory and chat search. Claude can draw on its memory and search your previous conversations; incognito chats are left out of both.6 Brello 1.0 carries only the last six messages of the current chat.
• Web search and web fetch. Claude can search the internet and read the full text of a page you link to.7
• Connectors. Web-based connectors such as Google Drive and Slack work on the web, desktop and mobile.8
• Agentic work and uploads. Anthropic’s policy covers chat, coding and “agentic sessions (where Claude performs multi-step tasks or takes actions on your behalf)”, and content you upload.1 Brello 1.0 takes text and at most one photo per message.

Where Brello is weaker

Brello 1.0 is a prototype with small on-device models, and it gives up convenience to keep processing on the phone.

• Availability. Brello 1.0 (version 1.0.0, 4 October 2026) is on Google Play and the App Store.
• Smaller models. On-device models are far smaller than frontier cloud models: they can be wrong, have a knowledge cutoff and are weaker at long or complex reasoning. Brello’s system prompt asks the model to admit uncertainty, which is no guarantee.
• Short memory. A 4,096-token context window; each answer sees only the last six messages, clipped.
• Hardware demands. Android 8.0 or newer on a 64-bit ARM phone, 12 GB of RAM recommended for Brello Pro, a 977 MB to 3.66 GB download plus a speed cache of similar size, and up to a minute for a model’s first launch.
• Fewer inputs. One photo per message; no files, no voice input or output, and an English-only interface.
• No sync, backup or export. If you lose or reset the phone, the chats are gone.
• Web search depends on others. Third-party engines can take up to about 14 seconds per attempt or return nothing, and they see your IP address.

Other comparisons

We answer the same questions for four other assistants, each from its provider’s own pages.

• Brello and ChatGPT: training, Temporary Chat and who sees your searches
• Brello and the Gemini app: activity storage, human review and Gemini Nano
• Brello and Microsoft Copilot: where your questions go
• Brello and Perplexity: two ways to answer from the web

Sources, with the dates we checked them

We opened each page ourselves on 5 October 2026. Where a page shows its own date, we give it too.

Facts about Brello describe Brello 1.0, version 1.0.0 (build 1), as of 5 October 2026. Product names are trademarks of their owners; Brello isn’t affiliated with them.