Meta’s Muse is an app. Muse Spark is the model. Don’t price them as one.

Meta spent 8 September 2026 introducing Muse as a personal agent that does the work, not just the answer. The durable builder object from that launch is a different name: Muse Spark 1.3, an API model. The app is what a person installs. The model is what a developer calls. This desk splits them on purpose. People searching for a Muse review want the app, which is graded separately. The model page is the Muse Spark profile. The token math is the pricing review. What follows is the product, then the sheet, then the part we will not pretend to have tested.

The app, as Meta described it

Muse rolls out in the United States on iOS, Android, and muse.ai, and inside WhatsApp. AI glasses are “coming soon,” which means they are not this launch. You tell it to send an email, book travel, fill a form, or push a longer goal. Meta says it keeps working after the app is closed and comes back for approval before it sends mail or pays. Checkout uses Link from Stripe, with a one-time card so the real card number stays hidden. Shop Pay and 1Password are named as later, not as shipping.

The free tier is usage-metered. Power is $20 a month and Maximum is $100 a month for more of that usage. TechCrunch reported that starting the app requires a payment card even when you expect to stay on the free tier, and that the app warns you before usage turns into a charge. Meta did not print the free allotment as a number of tasks. “Free for most of what people need” is a vendor sentence.

The security story, labeled

Meta says each person gets a Muse Secure VM: the agent and the connected credentials live there, a separate Sentinel must approve anything that reaches the internet, the model cannot see passwords or payment details, and there is an audit trail. People pick which apps connect and whether mail is read-only or send-capable. Conversations in that VM are not, Meta says, shared with its ad systems. You can tell it to forget something, and you can opt out of those chats training Meta’s models. A Confidential VM, encrypted to a key only the user holds so Meta itself cannot read it, is promised for later this year. Promised is not available.

Reuters reported the same week that internal tests had found Muse stalling and exposing sensitive data without authorization, and that the rollout proceeded with those concerns on the record. That is Reuters’ reporting of Meta’s own testing, not an AIToolsRatings run. A Sentinel diagram does not retire that report. If the reason you would install Muse is “it cannot leak,” read both the Meta post and the Reuters account before you connect a mailbox.

The model underneath

Muse Spark 1.3 is what Meta calls its most capable model for this kind of work, and it is also a public API preview for US developers. Standard price is $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15. A contributor checkpoint is $0.10 / $0.20 because Meta may train on those prompts and completions. Context is about a million tokens. Reasoning tokens bill as output.

Meta’s card compares 1.3 at max effort with GPT-5.6 Sol and Claude Opus 5, also at max, and with Muse Spark 1.2 at a lower effort. The flattering cells are long-context MRCR (98.5 and 98.1) and a slim DeepSWE lead (75.4 versus Opus 74.0 and Sol 73.0). The unflattering ones are GDPVal-AA, where Opus leads 1,824 to 1,754, and an internal instruction-following index where both Sol and Opus are higher. OSWorld and Terminal-Bench 2.1 are close. Terminal-Bench 2.1 is not the Terminal-Bench 4.0 column on our Grok page. The profile has the full grid and the effort caveat.

What we left out

• A directory listing. The app is US-only, and there is no global try-it URL that belongs in the tools catalog.
• A setup guide. This site does not onboard WhatsApp or API keys.
• A winner against Grok 4.7. The benches do not match, and we have not rerun either vendor’s sheet.

What to open next

If someone on the team says “Muse is $1.25,” send them to the pricing review so the $20 app plan and the contributor checkpoint do not get collapsed into that sentence. If they say “Muse beats Opus,” send them to the profile and the GDPVal row. If they want a typed decision instead of an agent that clicks a browser, this launch does not change the Jev page.