The Muse architecture
Muse runs on Meta's Muse Spark model. Meta Superintelligence Labs shipped Spark 1.1 in July 2026 with stronger reasoning and multi-agent coordination. The performance figures come from Meta's own published announcement and are vendor numbers, not independent measurements: 58% on Humanity's Last Exam in Contemplating Mode, and a Meta claim that the model matches Llama 4 Maverick at more than ten times lower compute. The same applies to the 1.3 revision, which Meta says needs 20% fewer tool calls and 25% fewer tokens than 1.2 — the kind of number that ends up on an API bill, but still a self-reported comparison.
Safety is handled by isolation, not restraint. Muse runs inside a cloud virtual machine called Muse Secure VM, and a second model, Sentinel, requires human confirmation before sensitive actions go through. Every agent vendor has landed on roughly this pattern: sandbox the model, then put a gate in front of anything that moves money.
The October privacy reports
Reports published in October 2026, including Android Authority's roundup of alternatives, described Muse reaching content it should not have touched and completing automated Facebook Marketplace sales for items their owners had not decided to sell. The same outlet ran a reader poll in which every respondent who said they would avoid Muse cited privacy as the reason.
A self-selecting poll is not evidence about the general public. It is evidence about who reads that site and cares enough to vote. The 100% figure should be read as sentiment among a specific audience, not as a measurement of consumer behaviour.
There is also the August 2026 incident that has not gone away. According to press reports at the time, a misconfigured test environment at the third-party auditor Irregular briefly allowed Muse Spark 1.1 uncontrolled internet access through a third-party vulnerability. On that account, Meta's own isolation design was not the failing component. The perimeter around it was.
Adoption and pricing
The adoption numbers are a mix of company statements and outside estimates, not audited figures. Meta says Muse hit number one on the US App Store shortly after release and reported 900,000 downloads in its first week, while third-party app-tracking estimates put it at roughly 3 million weekly and 1 million daily active users. Pricing follows the usual shape: a free entry tier, a $20 per month Power tier and a $100 per month Maximum tier. In euros, before VAT, that is roughly €18.50 and €92.60 at 1.08 USD per euro. European subscribers typically pay VAT on top, so the sticker price is not what lands on the card.
Europe is not on the launch list
Meta has launched Muse in the United States and Canada. No EU rollout has been confirmed. That is unusual for a Meta consumer product and it is worth separating the possible reasons. Agentic assistants need permissions that European data protection officers treat carefully: message access, calendar access, payment credentials, and continuous background execution. Under the GDPR, an EU company deploying such an agent is a controller for whatever the agent reads, and contracting with a vendor does not move that obligation. A "secure VM" is a technical claim, not a legal transfer.
There is also the calendar. The EU AI Act's core obligations for general-purpose AI models have applied since 2 August 2025, according to the European Commission's regulatory framework pages, with models already on the EU market before that date given until 2 August 2027 to fall in line. The AI Office's governance powers over general-purpose AI providers belong to that same August 2025 package; they are not a new August 2026 obligation. The later milestone is Article 50: its transparency duties apply from 2 August 2026, and machine-readable marking of generative systems already on the EU market before that date has a grace period until 2 December 2026. So this is no longer a preparation phase with voluntary guidelines. Providers of general-purpose models operating in the EU already face binding documentation and transparency duties, and an assistant that autonomously executes multi-step tasks across third-party services sits squarely in the area regulators have been asking questions about.
Alternatives and their euro prices
The alternatives named in the Android Authority piece are Claude Cowork, ChatGPT Work, Perplexity Comet, Vellum and MyClaw. Different tools, different assumptions about how much autonomy you want to hand over. What is comparable is the token cost behind them.
| Product | USD | EUR (at 1.08) |
|---|---|---|
| Meta Muse free / Power / Maximum | $0 / $20 / $100 per month | €0 / €18.50 / €92.60 before VAT |
| GPT-6 Luna API, input / output | $0.10 / $0.50 per 1M tokens | €0.09 / €0.46 per 1M tokens |
| Claude Haiku 5.5, input / output (under 100k tokens) | $0.10 / $0.50 per 1M tokens | €0.09 / €0.46 per 1M tokens |
| Mistral Large 4, input / output | $1.36 / $4.18 per 1M tokens | €1.26 / €3.87 per 1M tokens |
| Mistral Large 4, launch discount | $0.68 / $2.09 per 1M tokens | €0.63 / €1.94 per 1M tokens |
The spread is wide. For an agent that makes 20 to 40 tool calls per task, the difference between Haiku 5.5 rates and Mistral Large 4 rates is the difference between a cheap experiment and a budget line. Mistral is the one EU-headquartered option on that list, which matters if your procurement team asks where the data goes.
Agents you can host yourself
Open weights remain the only route to full data residency. GLM-5.3 ships under an MIT licence, and Llama 4 Maverick is available under the Llama 4 Community License. Both will run on a single consumer GPU at reduced context, and neither phones home with your calendar. We test local models through Ollama on an RTX 5060 Ti 16 GB and publish tokens per second, time to first token and VRAM figures on AI Arena, which is the fastest way to find out whether a given model fits your card before you download 40 GB.
What none of this solves is the approval layer. Sentinel-style confirmation is the part of Meta's design that other vendors are copying, and it is the part you have to build yourself if you self-host. A local model with shell access and no gate in front of it is a worse trade than a hosted agent with a privacy policy you dislike.
Is Meta Muse coming to Europe at all?
Meta announced and launched Muse in the United States and Canada. No EU launch date has been confirmed. Until one is, European users cannot sign up through normal app store channels, so the practical decision for most readers is which alternative to use rather than whether to switch.
How is an agentic assistant different from a chatbot?
A chatbot answers and stops. An agent plans a sequence of steps, calls external tools on your behalf, and keeps working after you close the window. Muse can hold a session across several apps and services, which is why it needs message, calendar and payment permissions that a chat interface does not.
Does the EU AI Act ban tools like Muse?
No. The AI Act does not prohibit agentic assistants. Its core obligations for general-purpose AI models have applied since 2 August 2025, and the Article 50 transparency duties apply from 2 August 2026, with a grace period until 2 December 2026 for machine-readable marking of generative systems already on the EU market before that date. These are transparency and governance duties, not prohibitions. A provider entering the EU market has to document how the model works and how it is supervised. For a company deploying an agent internally, the GDPR controller question is usually the more immediate constraint.
The comparison table above uses an assumed rate of 1.08 USD per euro and excludes VAT, which varies by member state. The alternative tools come from a single outlet's reader poll and a roundup, not from a controlled benchmark. Meta's Spark performance and adoption figures are the company's own claims, with third-party estimates where indicated.
