Meta's Muse family had a louder week than its marketing calendar probably planned for. Three things landed inside seven days: the model lineup showed up inside Google's enterprise storefront, the consumer agent started placing real phone calls for paying subscribers in the United States, and an unreleased version of the flagship model leaked through a public pull request.
None of those is the same story. Together they describe a company that has stopped treating Muse as a single product and started treating it as a portfolio with retail channels, consumer hooks, and a release treadmill.
Muse itself launched in September 2026 and climbed the US app charts quickly, passing 2.5 million downloads in its first few weeks. The model family sits under Meta Superintelligence Labs, with Alexandr Wang as the public face of both the models and the agent. Here is what each of last week's developments actually changes.
1. The Model Lineup Went Shopping for Storefronts
On November 9, Meta put three Muse models on Google Cloud through the Gemini Enterprise Agent Platform, and the same lineup is available on Oracle Cloud. That is a distribution story, not a capability story, and it is the one most likely to move enterprise revenue.
- Muse Ember (November 9): the smallest and cheapest Muse, built for high-volume agent loops. Roughly $0.10 per million input tokens, about $0.14 per task against $0.41 for Claude Haiku 5.5.
- Muse Spark 1.3 (September 2): the coding model, with tool calls down about 20% and token use down about 25% versus the prior Spark.
- Muse Wonder (October 12): the reasoning tier, priced for harder problems rather than throughput.
All three carry a one-million-token context window and accept text, images, video, audio, and PDFs. Ember also exposes 512,000 output tokens, which matters for agents that have to return long structured results rather than a paragraph.
The odd part is who is selling what. Google builds Gemini, Gemini competes with Muse directly, and yet Google's enterprise customers can now buy Muse through a platform carrying Gemini's name. That looks strange until you remember how cloud economics work: the platform earns when workloads run on its infrastructure, regardless of whose weights are doing the thinking.
The takeaway: Meta is not trying to win developers one signup at a time. It is renting shelf space from the two clouds that already hold enterprise contracts, which is a faster route to a procurement signature than any model card.
2. The Two Things Meta Didn't Announce Formally
Muse started dialing businesses. On November 12, Wang announced on X that paid US subscribers can send Muse to make outbound calls: restaurants, insurance providers, cable companies, the DMV, the usual hold-music purgatory. It waits on hold, navigates phone menus, and handles the conversation. The price of entry is the $20 monthly Muse subscription, with a $100 tier for heavier use.
The feature is not new in beta terms. Meta tested it in September 2026, and that test leaned on human contractors who stepped in when the model got stuck. Privacy complaints followed quickly, because a caller who believes software is on the line does not expect a stranger listening in, and Meta rolled that version back. The relaunch is positioned as AI-only with disclosure.
Competitively, Meta is late but not absent. Google's Call for Me expanded to Pixel 10 phones the same week, Grok Bot already calls consenting individuals, and Gemini is reportedly exploring calls to friends and family. Muse's differentiation is reach: it is bundled into a consumer app with millions of installs, not locked to one handset.
The leak came from a database tool. Pull request 15674 in Meta's public RocksDB repository added logging for an AI advisor used in an experimental live-tuning feature. The test notes recorded a call to a model identified as muse-spark-1.4-internal, with more than 33,000 input tokens, over 8,000 output tokens, and 7,867 of those classified as reasoning tokens.
That tells you a 1.4 checkpoint exists and is being exercised on real infrastructure work. It does not tell you the context window, the benchmark scores, the price, or whether the public name survives review. A single logged call is evidence of activity, not of progress.
The cadence is the more interesting signal. Spark 1.3 shipped September 2, Wonder followed October 12, Ember arrived November 9, and a 1.4 identifier is already appearing in internal tooling less than two months later. Meta is shipping on a six-week rhythm instead of an annual flagship cycle.
Three questions are worth tracking from here:
- Whether Ember's per-task pricing survives contact with real agent workloads, where retries quietly multiply the bill.
- Whether businesses push back on AI callers the way publishers pushed back on scrapers, since a receptionist's time is the cost being externalized.
- Whether Muse Spark 1.4 arrives as a Spark refresh or collapses the Spark, Wonder, and Ember tiers into one adaptive model.
The pattern across all three stories is the same. Meta is optimizing for surface area: more clouds, more consumer touchpoints, faster version churn. That is a reasonable strategy for a company that arrived late to the model race and cannot win on benchmark headlines alone.
It also makes the Muse brand harder to read. A subscriber paying $20 for phone calls and an enterprise buyer pulling Ember tokens through Google Cloud are using the same family name for two very different things, and Meta has not yet explained how the tiers relate.
Comments