# Meta Opens Its Muse Personal Agent to the Public With a Free Token Allowance

The agent runs in isolated virtual machines, reaches users through WhatsApp, and is built on the Muse Spark model line that scored 88.8 percent on Terminal-Bench 2.1 a week earlier.

- Published: 2026-09-09T06:20:57.202Z
- Canonical: https://polylog.news/ai/2026-09-09/meta-opens-its-muse-personal-agent-to-the-public-with-a-free
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Meta AI](https://ai.meta.com/muse/), [Meta AI](https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/)

Meta launched [Muse](https://ai.meta.com/muse/), a personal agent that books travel, files forms, handles email and shops on a user's behalf, on September 8. Each user's agent runs inside an isolated virtual machine with its own security boundary, the architectural answer to the obvious objection that an agent with credentials and browser access is a standing risk. Access is through a dedicated app and through WhatsApp, the distribution asset no competitor can match.

Pricing is unusually explicit about the underlying cost driver. Meta offers a free tier that Mark Zuckerberg described as usable up to about 100 million tokens a week, plus Power at 20 dollars a month and Maximum at 100 dollars a month, [according to reporting on the launch](https://www.marktechpost.com/2026/09/08/meta-introduces-muse-a-personal-ai-agent-that-runs-on-its-own-dedicated-secure-cloud-computer/). Metering consumers in tokens rather than messages shows that Meta expects agent sessions, not simple chats.

The model that powers Muse shipped a week earlier. Meta released Muse Spark 1.3 on September 2 into Muse Code and its [model application programming interface (API)](https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/), reporting 75.4 percent on DeepSWE 1.1, 88.8 percent on Terminal-Bench 2.1 and 98.5 percent on long-context retrieval. Those are vendor-reported numbers on agentic and coding suites, and Alexandr Wang, Meta's chief AI officer, told Axios the update was [groundwork for personal agents](https://www.axios.com/2026/09/02/meta-debuts-muse-spark-13-as-personal-agent-work-continues).

## What this means

Meta is competing on distribution and price rather than on frontier benchmark leadership, and a free tier measured in tokens is a direct subsidy against per-seat agent products from OpenAI, Anthropic and Google. The exposed group is the consumer agent startup layer, which buys inference at retail and now faces a rival that gives away a weekly token allowance large enough for most personal tasks. The second channel is inference demand: agents that browse, fill forms and retry failures consume far more tokens per user than chat, so Meta's own capital spending on serving capacity rises with adoption rather than falling.

## What to watch

- Whether Meta publishes retention or task-completion data for Muse, since agent products fail on reliability rather than on capability.
- Whether the isolated virtual machine design holds up against prompt injection through web pages and emails, the standard failure mode for agents with credentials.
- Whether rival consumer agents match the free token allowance, which would confirm that distribution and price, not benchmarks, decide this market.