PrivacyPick
In This Article
ComparisonsAI Chat

Claude vs Proton Lumo: Stronger Models or Encrypted Chats

Published November 14, 2026By Juan Martinez

Claude needs a phone number, stores chats Anthropic can read and trains on them unless you opt out. Lumo encrypts saved chats but has fewer models and tools.

claude-ai
VS
proton-lumo

Claude is the AI assistant from Anthropic, the San Francisco company that also builds the models behind it. Lumo is the AI assistant from Proton AG, the Swiss company behind Proton Mail, and Proton runs Lumo’s models on its own servers. Neither company sends your prompt to an outside model provider, so the comparison comes down to what each one keeps and what it can read.

This comparison covers the personal plans: what you hand over to start, who can read your chats, what trains the models, what each assistant can do, and what each costs. It does not compare answer quality, and the business plans of both companies run under separate terms.

ClaudeLumo
Rating5.37.5
Needed to startAccount and a mobile phone numberNothing; guest mode works without an account
Who can read saved chatsAnthropic; Trust & Safety staff open flagged chatsOnly you (zero-access encryption)
Trains on your chatsYes, unless you switch it offNo, except feedback you choose to send
How long chats are kept30 days after deletion; up to 5 years with training onGuest chats erased when the session ends; saved chats kept if you choose
JurisdictionUnited StatesSwitzerland
ModelsHaiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.2Lumo 2.0 Lite, Lumo 2.0 Max, Apertus 1.5
Paid plan$20/mo Pro, or $17/mo billed yearly$12.99/mo, or $9.99/mo billed yearly
AppsWeb, Windows, macOS, iPhone, Android; Linux in betaWeb, iPhone, Android

Want to compare every field side by side? Open the full comparison table.

Before the first message: a phone number or nothing

Claude does not work without an account, and every new account has to confirm a mobile phone number by text message. VoIP numbers, Google Voice and landlines are refused, the step cannot be skipped, and the number cannot be changed after verification. Anthropic says it uses the number only to stop spam and abuse. Accounts are open only to people who are physically in a country Anthropic supports. With the account, Anthropic also collects the IP address, the location worked out from it, and device and advertising identifiers.

Lumo can be used as a guest: you open the page and type, and the conversation is erased when the session ends. An account adds saved history that syncs between devices, and Projects. Proton says Lumo stores no metadata such as timestamps, IP addresses or chat context. No outside party has checked that statement.

Claude’s closest equivalent to a guest session is an incognito chat. It stays out of the history and memory and is not used for training, but it still needs the account and the phone number, and Anthropic keeps it for 30 days.

Who can read the chats you keep

Claude stores chats on Anthropic’s servers, encrypted in transit and at rest. That protects them from a stolen disk, not from Anthropic, which holds the keys (storage model). Staff cannot open conversations by default, but designated members of the Trust & Safety team can when a chat is flagged under the Usage Policy. Anthropic, PBC is a US company, so a US court order can reach the chats it stores.

Lumo stores saved chats under zero-access encryption. The keys are unlocked by your password. Proton holds nothing that would let it read saved chats, so a legal demand served on Proton AG in Switzerland returns data it cannot open. The trade-off is that Proton cannot unlock saved chats for you.

Training, and the feedback button on both sides

On Claude, one switch decides most of the rest: “Help improve our AI models” under Settings, Privacy (provider training claim). While it is on, Anthropic trains future models on your chats and may keep them in a de-identified form for up to five years. Anthropic’s pricing page lists model training as “Opt-out” on every personal plan, and when existing users were asked to choose in 2025, TechCrunch reported that the switch in the pop-up was already turned on. With it off, a deleted chat leaves Anthropic’s systems within 30 days, but chats already used in a training run stay in it.

Two paths lead to training even with the switch off. Chats that Anthropic’s safety classifiers flag can be used for its safety models, and are kept for up to two years. Rating an answer with the thumbs buttons sends the whole conversation to Anthropic as feedback, which it keeps for five years. Memory is also on by default: Claude saves facts from chats as topics you can review, edit or switch off.

Lumo does not train on chats, with one exception: pressing the feedback button can share an anonymised copy of that prompt and answer with the ETH Zurich and EPFL researchers behind Apertus 1.5. It happens only when you press the button. Proton’s Lumo pages say it never uses your data to train AI models and do not mention this exception, which appears on the support page.

Paying does not change Claude’s training default, storage or retention: they are the same from Free to Ultra.

What leaves around the chat

A May 2026 study by researchers at IMDEA Networks and UC3M in Spain found that, when a signed-in user accepts non-essential cookies, Claude’s website forwards the account email, plan and conversation ID to eleven advertising and analytics services. Rejecting those cookies stopped the forwarding, but the support-chat tool Intercom received the conversation’s address and the email whatever the cookie choice. The items are identifiers, not the text of the chat. Usage and log data are collected with no switch to stop them (telemetry). Anthropic has promised to keep Claude free of ads.

Lumo’s privacy policy says its phone apps may use app-store statistics and crash reports sent to a Sentry service that Proton hosts itself, and it documents no switch for that. When web search is on, only a simplified query goes to Proton’s search partners, not the whole prompt, and search can be turned off.

Models, tools and apps

Claude offers four Anthropic models from a menu (multi-model switching): Haiku 5.5 and Sonnet 5.5 on the free plan, plus Opus 5.5 on paid plans and the top model, Fable 5.2, which Pro pays for with usage credits. Every plan has web search, voice mode, artifacts, file creation with code execution, connectors to other apps and memory. Paid plans add Research, search across past chats, scheduled tasks, Claude Code and Claude in Chrome. File upload takes up to 20 files of 500 MB each per chat. Connectors let Claude read and act in services such as Google Drive and Slack, and each of those services handles what Claude sends it under its own privacy policy.

Lumo offers three models: Lumo 2.0 Lite and Lumo 2.0 Max, which run on open models that Proton adapts, and Apertus 1.5, built by ETH Zurich, EPFL and the Swiss National Supercomputing Centre. All three are on the free plan, with lower limits on Max, and Lite and Max each have Fast and Thinking modes. Image generation is on both plans, and a project holds up to 200 files, uploaded or linked from Proton Drive.

Lumo has no desktop app, and Proton’s Lumo page lists no voice mode and no connections to outside apps. Claude has apps for Windows, macOS, iPhone and Android, with a Linux app in beta. If your work depends on those tools, Lumo cannot replace Claude.

Open apps, unpublished routing, no audit

Claude’s apps and models are closed source (open source). Anthropic’s Trust Center lists a SOC 2 Type II report and ISO 27001 and ISO 42001 certificates, but the table of what they cover names the API and the business plans, not Free, Pro, Max or Ultra. For a personal account, every privacy statement rests on Anthropic’s own documents.

Lumo’s web, iPhone and Android apps are open source, so anyone can check what the apps send, and Proton’s Lumo page calls the code “fully open source”. The Lumo 2.0 models are not open in the same sense: Proton names the open models it uses without saying which one is behind Lite or Max. In 2025 the European Open Source AI Index found no published model routing, system prompt or architecture documentation and called Lumo the least open “open” AI assistant it had indexed. There is no independent security audit of Lumo, so the encryption design, the no-metadata statement and the no-training policy rest on Proton’s own description.

No breach of either service is on record.

Price

Claude’s free plan covers Sonnet and Haiku with web search, file uploads, memory, connectors and up to five projects, under a limit that resets every five hours and every week. Pro costs $20 a month, or $17 a month when $200 is paid for a year up front, and adds Opus, Research and Claude Code. Max costs $100 a month and Ultra $200 a month, both billed monthly. Prices exclude tax.

Lumo’s free plan includes every model and one project, but Proton does not publish its limits on messages, history, image generation or Max use. Lumo AI Plus costs $12.99 a month, or $9.99 a month billed yearly ($119.88), with a 30-day money-back guarantee.

Which one to pick

Claude fits if you need its range of models, Research, code execution, connected apps or desktop apps, and nothing you type is sensitive enough that Anthropic reading it would matter. Switch training off on the first day, leave the thumbs buttons alone, use incognito chats for one-off questions and reject non-essential cookies on the website. You still hand over a phone number, and Anthropic can still read what you keep.

Lumo fits if what you type matters more than what the assistant can do with it: a guest mode with no account, saved chats Proton cannot read, no training without your click, and a Swiss provider. In exchange you get three models, no desktop app, no voice mode or connected apps listed on Proton’s page, unpublished free limits, no audit, and Lumo-branded models whose weights and routing are not published.

Missing an app?

Tell us what you could not find. Every suggestion goes onto our list of products to review.

This opens a message in your email app, addressed to hello@privacypick.com. Nothing is sent until you press send there.