Muse Ember Release

Meta released Muse Ember on Nov 9, 2026, available that day to developers on Muse Code and the Meta Model API under the model name muse-ember-1.0. It costs $0.10 per million input tokens and $0.50 per million output tokens, the same list price Claude Haiku 5.5 starts from. Meta calls it its smallest and most efficient model, aimed at repeated agent work such as tool calls, coding subtasks, routing, extraction and parsing long documents. Meta's post says nothing about it reaching the Muse app or Meta AI.

What Meta says changed

All of this is Meta's own account, from its announcement:

• A small model trained by a big one. Ember is the first model Meta trained using Muse Wonder Full, its October frontier model. The post does not explain what that training involved.
• Four kinds of input in one model. It reads text, images, audio and video natively, without a separate model. The post describes no output beyond text.
• Built for agents. It supports the Responses API, and Meta says it tested whether the model respects user intent, handles sensitive information, resists hidden injected instructions and asks a person before high-impact actions.
• Safety review. Meta says Ember stayed below the high-risk thresholds for chemical and biological, cybersecurity and loss-of-control risks under its Superintelligence Scaling Framework.

What it costs

Source: Meta's announcement, read Nov 10, 2026. Meta links no separate price page, and one chart note in the post calls its cost figures estimates "at proposed pricing" while a later footnote says "list pricing", so these prices may yet be confirmed elsewhere.

Worked out from those prices, filling the whole 1 million token context costs $0.10, or $0.004 when all of it is cached, a 96% discount. The longest possible answer, 512,000 tokens, costs $0.26.

Since the list price matches Claude Haiku 5.5's, Meta's claim of more intelligence per dollar rests on how many tokens each model spends on a task, not on the rate. Meta puts one Artificial Analysis index task at about $0.14 on Ember against $2.39 on Muse Spark 1.3, for a score of 36 against 37.

Limits

The announcement gives a 1 million token context window and up to 512,000 output tokens. Meta has not stated rate limits, regions, a free tier, a knowledge cutoff or the full list of reasoning settings. Developers start at the Meta Model API or Muse Code.

Meta's benchmarks

Meta reports these at maximum reasoning effort in its announcement. Three of the nine tests are marked as Meta's internal evaluations.

Ember's one clear loss in Meta's own table is FrontierCode 2.0-AA, where Claude Haiku 5.5 leads by 7.7 points. Meta also lists wins on Agents' Last Exam (61.6%), OSWorld 2.1 (69.5%) and PerceptionBench (80.6%), its three internal tests, by 0.6 to 4.5 points over the next best model.

Meta posted no launch video for Ember on its Muse YouTube channel by Nov 10, 2026.

What creators say

1:20 Fahd Mirza, who reviews new AI models for 753K subscribers, gave Ember a sketch of a doner kebab stand in his Nov 10, 2026 test and asked for a working simulation. He says it answered in 41.1 seconds with two tool calls, kept the spinning meat cone, heat panel, knife and plate from the sketch, and added working heat, speed, slice and reset controls.

5:26 On a sales file he had made messy on purpose, he found the Excel report it built tidy: one date format, consistent names, discounts as percentages, a summary tab and charts in about 3 minutes, with rows missing a quantity "left blank instead of guessing", the test he says he set for hallucination.

7:08 Not everything held up. Asked to name a floating bookshop in 81 languages, it covered every language in about 7 seconds, but he says most names were plain translations, Bosnian and Serbo-Croatian came out identical, and a native speaker would need to check the rest.

What we think

We expect Meta to ship a second model trained with Muse Wonder Full before Mar 1, 2027. Meta presented Ember as the first of the specialized models it said Wonder would help build, and Muse Spaces and the voice avatars show how quickly it is adding to Muse this autumn.

Confidence: Likely. Check by: Mar 2027.

We think Ember reaches the Muse app as its fast option within six months. It is priced for high volume and takes text, images, audio and video, though Meta's post so far names only developer products.

Confidence: Possible. Check by: May 2027.

Sources

• Introducing Our Smallest, Most Efficient Model: Muse Ember, Meta, Nov 9, 2026, read Nov 10, 2026
• Meta Model API and Muse Code, Meta, read Nov 10, 2026
• Meta Muse Ember: Tiny Cost, Big Test on Images, Excel and Circuits, Fahd Mirza, Nov 10, 2026 (video)