Muse Agent Is Live: Meta's Consumer AI Agent Runs on Muse Spark 1.3

Meta switched on Muse Agent — the consumer AI agent it has been hinting at all year, branded simply "Muse" inside its own new iOS and Android apps — in the United States on September 8, 2026. It is powered by Muse Spark 1.3, the same checkpoint Meta put on its developer API exactly one week earlier, and that lineage is the real story: Muse Agent is the productized form of a model whose release notes were written about long-horizon agentic work. Meta describes the agent as running on its own computer — a cloud virtual machine with a file system and a terminal — so it can browse the web, write its own code, build small tools, and keep working on tasks after the app is closed.

This is not another chatbot skin. Muse connects to a person's email, calendar, payment, and shopping accounts and acts on them: send email, book travel, buy things, negotiate, fill out forms. Meta says it deliberately delayed the product from an internal April target to harden it after security testing found real flaws, and the launch is US-only for accounts aged 18 and over, and geo-gated — the profile of a company keeping the first version of an agent with spending power inside a blast radius it can still control. What follows covers what Muse Agent actually does, the Muse Spark 1.3 engine underneath it, the security architecture Meta built to justify giving an AI a wallet, and what the pricing says to a developer building an agent of their own.

The pitch: an agent with its own computer

Meta's design post for the product states it plainly: Muse has its own computer, with a file system and a terminal, which it can use to write code and build tools — alongside a full browser for navigating sites, filling forms, and completing transactions. That is an architectural statement rather than marketing flourish: Muse is not a large language model answering from context, it is a model operating a machine, the way Muse Code operates a developer's machine. Long-running goals, such as planning a year of training or standing up a small business, are decomposed into action plans the agent works through over time, and it can run on a schedule or in response to events, notifying the user only when something is genuinely new or needs a decision.

The interaction model is a departure from assistant chat too: one long, interruptible main conversation that holds context across days, separate side chats for distinct tasks, persistent memory files the user can read and edit, and "artifacts" — documents, PDFs, web pages, dashboards, grocery lists — that can live outside the chat. Users can name the agent and give it an avatar and a communication style, which launch coverage treats as a core reason people bond with it. On integrations, Meta lists Google Workspace (Gmail, Calendar, Drive), Ticketmaster, OpenTable, Spotify, and Apple Health at launch, with Shopify's Shop Pay and 1Password planned, plus custom API connectors users can generate for services without a built-in integration.

The engine: Muse Spark 1.3, one week off a developer launch

Underneath is Muse Spark 1.3, which Meta released on September 2 through Muse Code and the Meta Model API. It is the fourth checkpoint in the Muse Spark line since April and, on the vendor's own figures, the largest step in coding and agentic work the family has taken: roughly 20% fewer tool calls and about 25% fewer tokens than Muse Spark 1.2 on comparable tasks — the kind of efficiency that matters when a single agentic task can span thousands of calls. The standard-tier price was held exactly at the Muse Spark 1.2 level — $1.25 per million input tokens and $4.25 per million output — with a $0.10/$0.20 Contributor tier for users who let Meta train on their traffic, and the context window stays at 1 million tokens.

The independent picture is starting to fill in. On the Artificial Analysis Coding Agent Index, which pairs each model with its native agent harness, Muse Spark 1.3 running in the Muse Code harness scores 68 at the top "max" reasoning setting — second only to Claude Opus 5 (xhigh) in Claude Code, and ahead of Claude Fable 5 (max) — while the shipping xhigh configuration scores 64. A scope note on that result: it measures the model driving its own coding agent, not the consumer Muse Agent product, and Meta has not said which reasoning configuration the consumer agent runs. Meta's own launch numbers — Terminal-Bench 2.1 at 88.8, DeepSWE v1.1 at 75.4 at max, OSWorld 2.0 at 66.9 at max — remain vendor-reported and independently unreproduced at this writing.

The consumer and developer threads connect here. The agent Meta sells for a monthly subscription is the product form of a model a developer can call for $1.25 per million input tokens, and the same model that produced its strongest coding results inside Meta's own Muse Code harness. Muse Agent is Meta taking its frontier model to the end user directly, before third parties have had time to build the same product on top of the API — which is precisely the position its Superintelligence Labs was created to defend.

The security architecture is the product

The reason anyone hesitates to give an AI access to email, calendars, and a payment method is exactly where Meta spent its delay. Each Muse instance runs in a dedicated Muse Secure VM — a cloud virtual machine that isolates the agent and the user's data from the open internet. A separate system-level agent called Sentinel is the only path out: every interaction with the outside world passes through it, it holds the policy, and it cannot be overridden by Muse itself. Sensitive actions — sending an email, spending money — surface approval cards the human accepts or rejects, and the VM keeps an activity log under the user's control.

The credential and payment design follows the same principle. Passwords live in a secure credential store the agent cannot read, and purchases are executed through Stripe's Link infrastructure, which hands the agent a single-use card number per transaction rather than the user's real card. Meta says conversation data is not shared with its advertising systems and that users can opt out of training use. The roadmap item to watch is the planned Muse Confidential VM, developed with Signal creator Moxie Marlinspike, in which the user — not Meta — holds the cryptographic keys, meaning even Meta could not open the VM. That would be a first for a mass-market consumer agent, and it is not shipping today.

Pricing: subscriptions for people, tokens for developers

Muse's consumer pricing is a three-tier subscription rather than usage billing. The free tier runs on a roughly 100-million-token weekly allowance with a weekly reset; Meta says most people should stay on the free tier and shows a usage meter before an action crosses into paid territory, though a payment card is required on file even for free — a friction several reviewers flagged. Paid tiers run $20/month and $100/month, dubbed Power and Maximum in launch coverage, for users who exceed the free allowance or want continuous high-compute workloads.

If what you want is to build an agent with this behavior rather than rent Meta's, the interesting number is the per-token one, and the model economics here are unusually legible. Muse Spark 1.3 is available to any developer through the Meta Model API at $1.25 per million input tokens and $4.25 per million output, with reasoning tokens billed as output and cached input at $0.15. The 25%-fewer-tokens claim is the part to watch, because at an unchanged list price it is the entire cost story. On routing and failover, OrcaRouter's position is the one we have stated on every Muse Spark 1.3 piece: the new checkpoint is not yet in our directory as of this writing on September 9 — we do not list what we cannot serve. Its predecessor is: Muse Spark 1.2 is live on OrcaRouter at Meta's exact list price, $1.25 in and $4.25 out per million tokens with no markup, so the "price held" claim above is checkable against a real page. Because we pass provider list prices through rather than adding a margin, the day a routed provider starts serving Muse Spark 1.3 it appears here at Meta's price the same day, with automatic failover across providers for production agent workloads.

What to watch

The most important open questions are not on a benchmark sheet. First, whether the consumer product's reasoning configuration and safety boundaries hold up in the wild — the April delay was reportedly prompted by real security-testing failures, and an agent that can spend money is held to a different standard than one that can write code. Second, the 100-million-token free allowance is a statement about agentic economics: if Meta can run a genuinely useful personal agent inside that envelope at scale, subscription pricing becomes a serious answer to how agent capability gets sold, and the $1.25/$4.25 API rate starts to look like the floor for everyone else building the same thing. Third, the Confidential VM, the promised open-weights release for the Muse Spark family, and the rollout of the underlying model into Instagram, Facebook, and Meta AI apps over the coming weeks. Muse Agent is day one of that sequence, not the end of it.