Table of Contents
Lumo 2.0 vs ChatGPT vs Claude.ai: the private AI assistant for teams that cannot share their data (July 2026)
Proton shipped Lumo 2.0 on June 30, 2026, and for the first time its private AI assistant posts a score worth arguing about. On the Artificial Analysis Intelligence Index, Lumo 2.0 Max lands at 51, versus 59 for GPT-5.6 Sol and 60 for Claude Fable 5 (Artificial Analysis, July 2026 leaderboard; Lumo's 51 confirmed on Proton's launch post). That eight-to-nine-point gap is the entire decision.
Here is the short version. If your team touches data you are not legally allowed to hand to a US AI vendor (patient records under HIPAA, client files under GDPR, anything covered by CCPA), Lumo 2.0 is now capable enough that the privacy tradeoff pays for itself on most daily work: drafting, summaries, research, everyday code. Pick Lumo for Business. If you have no data-sharing restriction and your team lives in complex code generation or long-horizon reasoning, the higher ceiling on ChatGPT or Claude is worth keeping. The gap is real and it shows up exactly where the hard work is. The rest of this piece is where that fork flips, task by task, priced and dated.
What Lumo 2.0 actually changed
The old objection to Lumo was simple: it was private, and it was noticeably weaker than ChatGPT. Version 1.4 scored 15 on the same Intelligence Index. Lumo 2.0 Max scores 51, a 240% jump, and the lighter Lumo 2.0 Lite scores 34, up 127% (Proton). That is the difference between "privacy protest vote" and "credible daily driver."
Five concrete changes drive it, all documented on the launch post:
- Two reasoning modes. Fast prioritizes speed; Thinking runs multi-step reasoning with a visible thinking state. Proton says everyday queries return up to 76% faster than Lumo 1.4 (Proton).
- Multimodal image support. You can upload an image to analyze a chart or screenshot, generate visuals from a prompt, or edit an existing image, all inside one encrypted conversation.
- Live web search with source citations. Earlier versions leaned on the model's own knowledge; 2.0 pulls current results and links them so you can verify the answer.
- Memory plus a context window that is now twice as large, alongside encrypted Projects that keep chats, files, and instructions together across sessions.
- Custom Lumos, which are saved assistants with fixed instructions you write once.
That last one is the most useful feature nobody demos well. A Custom Lumo is a reusable instruction set, so a compliance team can pin its house rules once instead of re-pasting them every session:
Name: Contract redline assistant
Instructions:
- Draft in plain US English, no legalese unless quoting a clause.
- Flag any clause that shifts liability to us; quote it verbatim.
- Never invent a statute or case cite. If unsure, say "verify with counsel."
- Output: 1) summary, 2) redline table, 3) open questions.Andy Yen, Proton's CEO and a former CERN physicist, describes the release as "re-engineered from the ground up," and argues its Thinking mode is no longer perceptibly different from the latest OpenAI and Anthropic models for many everyday tasks (IntelligentHQ, July 1, 2026). That is a vendor claim, and the benchmark keeps it honest: "many everyday tasks," not all of them.
The capability gap, task by task
The single index score hides where Lumo competes and where it does not. Here is the ranking that matters, on the same benchmark:
| Model | Artificial Analysis Intelligence Index | Source |
|---|---|---|
| Claude Fable 5 | 60 | Artificial Analysis |
| GPT-5.6 Sol (max) | 59 | Artificial Analysis |
| Grok 4.5 (high) | 54 | Artificial Analysis |
| Lumo 2.0 Max | 51 | Proton |
| Gemini 3.5 Flash | 50 | Artificial Analysis |
So Lumo 2.0 Max sits mid-pack among frontier reasoning models, ahead of Gemini 3.5 Flash and a few points behind Grok, with an eight-to-nine-point gap to the two leaders (Artificial Analysis, July 2026). Points on a composite index do not spread evenly across real work, though. Break it down:
Document summarization and editing. This is the smallest gap. A May 2026 comparison of the pre-2.0 Lumo already found it "competent and generally fast" on summaries, email drafting, and file analysis, matching ChatGPT on basic productivity tasks (factually.co, researched May 2, 2026). With 2.0's larger context window and reasoning modes, this is where a compliance team loses the least by switching.
Web research. Formerly a weak spot. Lumo 2.0's live search now returns current results with citations, which is the exact shape of answer a researcher can check (Proton). That same May review flagged the older web search as unreliable, so treat this as the feature that moved the most and test it against your own queries before you trust it fully.
Coding. This is the widest gap, and the honest reason to keep a frontier model. That earlier comparison found Lumo's outputs "more generic, less nuanced, and less well-structured than ChatGPT," and traced it to the relatively small open models Proton runs for privacy and cost reasons (factually.co). Version 2.0 narrows this, but 51 against 59 to 60 is exactly the distance you feel on multi-file refactors and dense algorithmic work.
Creative and long-form writing. Solid for everyday output, weaker at the top end. Reviewers found Lumo's prose less varied than ChatGPT's on nuanced creative tasks (factually.co). For a marketing draft or an internal memo, fine. For a flagship piece where voice carries the work, the frontier models still edge it out.
The candid con, stated plainly: Lumo runs smaller open-source models than its US rivals, and that is why the coding gap is real rather than cosmetic. Proton trades some raw capability for the ability to run everything on infrastructure it controls. For daily knowledge work, you rarely hit the ceiling. For heavy engineering, you hit it by Thursday.
What the privacy guarantee actually buys you
This is the reason to be here, so be precise about it. Lumo's privacy is architectural, not a promise in a terms-of-service PDF.
Every Lumo conversation is zero-access encrypted, and the same architecture secures over 100 million Proton Mail and Proton Drive accounts (Proton). Proton's design means Proton itself cannot read your conversations, and nothing you type is used to train a future model (IntelligentHQ). Lumo is also fully open source, so the encryption claims are inspectable rather than taken on faith (ItsFOSS, June 30, 2026).
Where it runs matters as much as how it is encrypted. Lumo operates on independent European infrastructure under Swiss privacy law, which means access to it is not subject to US Executive Orders and user data is not exposed to American data collection requests (Proton). For an EU-headquartered company or a US firm nervous about jurisdiction, that is the line item a general counsel actually reads.
Now the factual contrast, without the alarm. On ChatGPT's consumer tiers, conversations can be used to train models unless you turn that setting off, and memory features store data by design; check the current data controls on OpenAI's pricing and settings. On Claude.ai, Anthropic's policies allow access to conversations for safety review under defined conditions (Anthropic / Claude pricing). Both companies offer business and enterprise tiers with stronger data commitments, and for many teams those are enough. The difference is the shape of the guarantee: a contractual "we will not" versus an architectural "we cannot." If your compliance posture has to survive an audit rather than a trust exercise, that distinction is the product.
Pricing, dated July 13, 2026
Prices below are current as of July 13, 2026. Competitor tiers are billed in the vendor's own currency and change often, so treat the business figures as a range and confirm at the linked page before you buy.
| Tier | Lumo | ChatGPT | Claude.ai |
|---|---|---|---|
| Free | Yes, limited Max model usage and history (Proton) | Yes, limited (OpenAI) | Yes, limited (Anthropic) |
| Personal paid | Lumo Plus, $9.99/mo billed annually ($12.99 month to month) (Proton) | Plus, about $20/mo (OpenAI) | Pro, about $20/mo (Anthropic) |
| Business | Lumo for Business, $11.99/user/mo billed annually (Proton) | Business, roughly $25-30/user/mo (OpenAI) | Team, roughly $25-30/seat/mo (Anthropic) |
| Verified compliance posture | GDPR, HIPAA, CCPA plus zero-access encryption (Proton) | Contractual, per plan (OpenAI) | Contractual, per plan (Anthropic) |
One line changes the math for teams already thinking about privacy. Proton Unlimited bundles Lumo Plus with Proton Mail, Drive, VPN, and Pass in a single subscription (see Proton's plan pricing). If you are pricing a private email plus VPN plus password manager plus an AI assistant, the bundle is a different comparison than Lumo Plus on its own, and it is the number the Proton-ecosystem team should run.
Who should pick what
Match the row to your constraint, then read the paragraph under it.
| Your situation | The pick | What flips it |
|---|---|---|
| Regulated data (healthcare, legal, finance, EU HQ) | Lumo for Business | If a workflow is 80% heavy code, keep a frontier model beside it |
| High-output dev or research team, no data restriction | ChatGPT or Claude | If a client contract bans US data processing, Lumo moves in |
| Already paying for Proton | Lumo Plus inside Proton Unlimited | If you need HIPAA admin controls, upgrade to Lumo for Business |
Privacy-first compliance team. Lumo for Business is the pick, and the reason is the architecture, not the marketing. The zero-access design means you are not relying on a vendor promise that your patient notes or client files stay unread; the system is built so they cannot be read (Proton). Start on Lumo Plus to sanity-check capability on your real workload for a week, then move to Business for the admin controls and the GDPR, HIPAA, and CCPA coverage. The candid con: your engineers will notice the coding gap, so pair it with a frontier model for the pure-code lane and keep the regulated work inside Lumo.
High-output dev or research team without compliance constraints. Stay on ChatGPT or Claude. At 59 and 60 on the index against Lumo's 51, the leaders pull ahead exactly where you spend your day: multi-file code, dense reasoning, top-end creative work (Artificial Analysis). The smart hybrid is to use Lumo as the private lane for anything sensitive (an unreleased roadmap, a legal question, a customer's data) and keep the frontier model for the hard build. You are buying the ceiling, and you would feel it if you gave it up.
Proton-ecosystem team. If you already run Proton Mail and VPN, Lumo Plus is essentially free inside Proton Unlimited, which folds it in with Proton Mail, Drive, Proton VPN, and Proton Pass under one bill. Compare that bundle to the cost of a separate email tool plus a standalone AI subscription and the arithmetic usually lands in Proton's favor. The flip: the moment you need HIPAA-grade admin controls and audit logs, step up to Lumo for Business rather than stretching the personal bundle across a team.
The verdict
For a team with real compliance exposure, Lumo 2.0 is finally the answer instead of the compromise: pick Lumo for Business, and accept a single, known tradeoff on heavy coding. For an unconstrained team chasing the capability ceiling, ChatGPT or Claude still win on the work that is actually hard, and Lumo earns a spot as the private lane for sensitive prompts. For anyone already inside Proton, the assistant comes bundled, and Proton Unlimited is where you start.
You can also evaluate the assistant on its own, no ecosystem required, at lumo.proton.me. Run your own hardest daily task through it for a week. The benchmark says 51 against 59 to 60; whether that gap matters is a question only your actual workload answers, and now it is close enough to be worth asking.