Meta Superintelligence Labs introduced Muse Ember on Nov. 913. Small is the point. It is the fourth model under the Muse brand, and the smallest one yet.
The new model handles the small jobs agents repeat all day. It calls tools, fixes code, sorts documents, and pulls records1. Ember also reads screenshots, charts, scanned pages, voice memos, and video. Its context window holds a million tokens at once.
Wonder and Ember work as a pair on harder jobs. One plans the approach, and the other carries out each repeated step. Meta put it this way in its launch post. Ember is "our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day"1.
Chief executive Mark Zuckerberg posted the launch on X that same afternoon24. Alexandr Wang is Meta's chief AI officer. He called the new model "smarter than luna" in a post of his own, and cheaper than Haiku too.
A Tied Score, at the Top Price
Meta built one chart to make its case. On the Artificial Analysis Intelligence Index, Ember scores 36 at its hardest reasoning setting. That lands almost level with Claude Haiku 5.5's 36.2, and well ahead of GPT-6 Luna's 281. Call it a tie.
Muse Ember Ties Claude Haiku's Intelligence Score at 36
Score on the Artificial Analysis Intelligence Index, each model at its top reasoning setting; higher is better
Each column is one model's score on the index Meta cites in its own launch post; higher means a stronger score. Ember and Haiku land almost level, and Luna sits well below both.
Meta ran and scored this comparison itself, and only at each model's most expensive reasoning setting.
Source: Meta [1]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Muse Ember | 36 |
| Claude Haiku 5.5 | 36.2 |
| GPT-6 Luna | 28 |
Price sits beside that score. Meta puts Ember's cost near 14 cents per task at that same setting14. Haiku costs 41 cents, and Luna costs 15 cents.
Ember undercuts the model it ties on intelligence by close to two-thirds. That gap is real money, multiplied across millions of agent tasks a week.
Muse Ember Costs 14 Cents a Task, a Third of Haiku's 41
Meta's estimated cost per task at matched reasoning effort, in dollars
Each column is the dollar cost Meta assigns to one task at its model's top setting; Ember's column is the shortest of the three.
The gap holds only at the top setting Meta compared; Haiku offers cheaper, lower-effort settings this chart leaves out.
Sources: Meta [1]; The Next Web [4]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Muse Ember | $0.14 |
| GPT-6 Luna | $0.15 |
| Claude Haiku 5.5 | $0.41 |
The comparison carries a catch. Haiku offers cheaper settings, and Meta's chart skips them4. The saving applies only when both models run at full effort.
Meta also wins the chart with its own evaluations. AutomationBench is one: Ember scores 37.8%, Luna scores 20.2%, and Haiku scores 13.9%1.
One test still favors Haiku. Haiku keeps that edge on FrontierCode 2.0, a coding benchmark, scoring 34.9% against Ember's 27.2%1. The Next Web's Ana Maria Constantin summed up the mixed picture plainly. Ember "beats GPT-6 Luna on Meta's tests but trails Claude Haiku on one coding benchmark," she wrote4.
Five Companies Tested It Early
Meta built Ember for agent harnesses, and early testers back that pitch with their own numbers. The numbers hold up. Yashodha Bhavnani is Box's vice president of AI products. She said Ember showed "impressive accuracy, speed and token efficiency" scoring contract clauses line by line1.
Saoud Rizwan founded the coding agent Cline, and he called Ember "purpose-built for" subagent work like coding and document review.
Dane Knecht is Cloudflare's chief technology officer. He reported Ember ran "up to 30% faster at the tail" on multi-repo code investigation1. Konstantinos Papamiltiadis holds a similar rank at Oracle, serving as senior vice president of global partnerships, the role Meta's blog post lists for him. He said Ember showed "an encouraging ability to tackle complex engineering tasks" in early testing.
Chak Pothina leads inference at You.com. Speed sold him. He said Ember finished every coding task in the company's own benchmark with a 2.2-second median response. It also used 36% to 48% fewer output tokens than the models it replaced.
AI at Meta@AIatMetaToday we're introducing Muse Ember, our smallest and most efficient model, built for complex reasoning execution and everyday agent work.
Muse Ember can carry workflows across apps, call tools in sequence, run tests and make targeted code fixes. With a 1M-token context window and native understanding of text, images, audio and video, it can also work across large document sets and varied inputs.
Four Models Deep Into the Muse Line
Ember follows three earlier Muse releases. Four models in seven months.
Meta shipped Muse Spark in April3. It shipped an open twin, Muse Glimmer, in August. The larger Muse Wonder followed in October.
Muse Agent is Meta's consumer product built on the family, and it launched in September. New features, including shared Spaces and video chat, arrived earlier this month.
Pricing stays steady across the line's cheapest tier. Ember costs 10 cents per million input tokens and 50 cents per million output tokens1. Cached input costs 0.4 cents per million. A prompt past 200,000 tokens steps up to 30 cents per million input and $1.50 per million output.
Meta says Ember also undercuts open rivals DeepSeek V4.1 Flash and GLM 5.3 Flash on price. Zhipu AI's Z.ai builds that second model3.
Ember runs today as muse-ember-1.0. It lives on Meta's own developer platform and inside Muse Code, its coding tool1.
Google Cloud and Oracle Cloud added Ember this week, alongside Muse Spark 1.3 and Muse Wonder. The timing looks deliberate. That gives buyers who already hold a cloud contract a direct path to Meta's models5.
Meta also ran Ember through its Superintelligence Scaling Framework, the company's own test for chemical, biological, cyber and loss-of-control risk1. The model stayed below every high-risk threshold, Meta reported.
Its agent-behavior testing covers one more worry. It checks whether Ember resists hidden instructions buried inside web pages, documents or another tool's output. That worry grows as agents read more of the open web on a person's behalf.
What to Do
Read Meta's own chart before trusting the headline number. Context is cheap. The tie with Haiku holds at one setting, the costliest one1. Haiku still leads on at least one named test4.
Weigh Meta's internal evaluations apart from the outside ones. AutomationBench and the Personal Superintelligence Index come from Meta's own testing floor1. The Artificial Analysis Intelligence Index, by contrast, runs on a chart a third party built.
By the Numbers
- 36 and 36.2: Ember and Haiku 5.5 tie on Artificial Analysis's Intelligence Index, each at its top setting1.
- 28: GPT-6 Luna's score on that same index, lowest of the three1.
- 14 cents, 15 cents and 41 cents: Meta's estimated task cost for Ember, Luna and Haiku at that top setting14.
- $0.10 and $0.50: Ember's price per million input and output tokens1.
- 37.8%: Ember's score on Meta's AutomationBench, ahead of Luna's 20.2% and Haiku's 13.9%1.
- 34.9%: Haiku's score on FrontierCode 2.0, a coding test, ahead of Ember's 27.2%, the one named benchmark Haiku still leads1.
- 1 million: Ember's token context window, read in a single pass1.
- Four: Meta's total model count under the Muse brand since April3.
What to Watch
Watch for an outside lab to publish its own Ember score. So far, every benchmark stays in-house. Each number traces back to Meta's own test runs, or to the Artificial Analysis chart Meta chose to show1.
Anthropic's response is worth watching too. Rivals rarely sit still. Haiku 5.5 still carries a cheaper settings tier, and Meta's comparison leaves that tier out4. A price move from Anthropic would close the gap Meta just drew.
Sources
- Meta, "Introducing Our Smallest, Most Efficient Model: Muse Ember," Meta, Nov. 9, 2026, https://research.meta.ai/blog/muse-ember
- AI at Meta (@AIatMeta), "Today we're introducing Muse Ember," X, Nov. 9, 2026, https://x.com/AIatMeta/status/2119853591703998787
- Russell Brandom, "Meta releases Muse Ember, a new lightweight model for agent execution," TechCrunch, Nov. 9, 2026, https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/
- Ana Maria Constantin, "Meta's Muse Ember matches Claude Haiku at a third of the cost per task," The Next Web, Nov. 9, 2026, https://thenextweb.com/news/meta-muse-ember-small-ai-model-haiku-luna
- Diego Almada Lopez, "Meta's Muse Ember, Muse Spark 1.3, and Muse Wonder land on Google Cloud," Crypto Briefing, Nov. 9, 2026, https://cryptobriefing.com/meta-muse-models-google-cloud/
