On November 9, 2026, Meta Superintelligence Labs released Muse Ember — the fourth model in the Muse family, and the smallest and cheapest of them. The launch post's title states the positioning outright: "Introducing Our Smallest, Most Efficient Model: Muse Ember". The same day, Anthropic released Claude Fable 5.2 / Mythos 5.2, putting the industry's biggest and smallest models on stage on the same day — and the coverage plays on exactly that contrast. This article collects everything confirmed on launch day: the timeline, the family position, the training background, the five launch channels and the first wave of coverage.
Ember was not a teaser-then-API model: on November 9 the Meta research blog post went live, Alexandr Wang announced it on X the same day, model ID muse-ember-1.0 entered the Meta Model API, Muse Code made it available on all plans, and Google Cloud, Oracle Cloud and the Vercel AI Gateway all connected the same day. Artificial Analysis published its launch evaluation the same day (Intelligence Index 36, $0.14 per task); TechCrunch covered it at 18:00 UTC; the Hacker News launch thread reached 228 points that day. Launch, availability, independent evaluation and mainstream coverage all landed on the same day — the clearest difference between Ember and the many models that ship a paper weeks before an API.
The Muse family runs to four models with Ember: Spark (April, mid-weight mainstay), Glimmer (August, the open-weight twin of Spark), Wonder (October, frontier flagship), Ember (November, smallest). Ember's role is not a shrunken flagship but the executor in a division of labour — Meta's own words are "Muse Wonder plans, Muse Ember handles the volume": Wonder plans, Ember takes the high-volume execution. The launch also set the family's internal price anchor: Ember's AA Index of 36 sits one point below Spark 1.3's 37, at roughly a seventeenth of the cost per task ($0.14 vs $2.39).
Alexandr Wang stressed on launch day that Ember is the first model Meta has released trained using Muse Wonder Full. Wonder Full is the largest Wonder variant, locked inside Muse Code and not offered externally. The training method itself was not disclosed, and the parameter count was not published — both are explicitly missing from Ember's launch post, and no other model's numbers should be used to fill them in. Meta's post also notes Ember was safety-evaluated under the Superintelligence Scaling Framework (chem/bio, cyber, loss-of-control), with no category exceeding the high-risk threshold; it was also evaluated for resisting instructions hidden in web pages, documents and tool outputs, and for seeking human input before high-impact actions.
Meta's launch post quotes five partners: Box, Cline (founder Saoud Rizwan), Oracle, Cloudflare and You.com. These are Meta-published quotes, not independent third-party verification — cite them as such. Oracle is also one of Ember's launch-day channels.
On the same November 9, Anthropic released Claude Fable 5.2 / Mythos 5.2, AA Index 65, No. 1 on the board. i10x put it plainly: Fable grabbed the frontier headlines. Ember did not compete on peak capability — its headline is a different race: AA Index 36, tied with Haiku 5.5, at a third of the cost per task, the "cheapest model scoring 36 or higher". This site already has a Fable 5.2 topic, and the two read as a pair: one about the biggest, one about the cheapest. agentfred published a same-day comparison of Fable / Mythos against Ember on Nov 10–11, the only analysis so far to put both in one piece.
English mainstream coverage came from TechCrunch (18:00 UTC on the day) and The Verge (same-day coverage; this site did not obtain the original link in this round, so it is not linked here); The Decoder was cited by TW/HK media as the source of the "more intelligence per dollar" claim. In Chinese: datalearner followed up on Nov 9 itself, cocoloop and jiazhuangai on Nov 10, and gptclub published a detailed report on Nov 11 ("same price as Haiku, a third of the cost per task"); TW/HK added techbang, sunmedia and airesearch.hk. But no piece appeared from 机器之心, 新智元, 量子位 or IT之家, and Chinese autocomplete is zero — for now, Chinese readers learn about Ember mainly through the English launch and niche reposts.
Yes. The Meta research blog post, Alexandr Wang's X announcement, the Meta Model API listing, Muse Code on all plans, Google Cloud, Oracle Cloud, the Vercel AI Gateway and the AA launch evaluation all happened on November 9, 2026. There is no date ambiguity.
Yes. Anthropic's Claude Fable 5.2 / Mythos 5.2 also released on November 9. This site already has a Fable 5.2 topic — AA Index 65, No. 1 on the board; Ember's AA Index is 36, and its story is value, not peak capability.
The fourth. Spark (April), Glimmer (August, open-weight twin), Wonder (October, frontier flagship), then Ember (November, smallest).
Yes. The Meta Model API (muse-ember-1.0), Muse Code on all plans, Google Cloud, Oracle Cloud and the Vercel AI Gateway were all ready on launch day. There is no free tier and no downloadable open weights — trying it means going through an API.
Reference: Meta research blog — Introducing Our Smallest, Most Efficient Model: Muse Ember · Artificial Analysis — launch evaluation (Index 36, $0.14 per task) · TechCrunch (Nov 9, 2026, 18:00 UTC) · Hacker News launch thread (228 points) (rewritten and organized by ModelBrief; facts per the Meta launch post and the original reporting)