ModelBrief · Muse Ember · Topic Page
Last updated Nov 20, 2026

Muse Ember vs GPT-6 Luna and the Muse Family: Planner and Executor

ModelBrief · Published 2026-11-16 · Fact-check 2026-11-16 · Reference: Meta research blog, Artificial Analysis

GPT-6 Luna is the other rival Meta names in its launch post — alongside Haiku 5.5 in the headline claim ("more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna"). But the Luna comparison has a quirk: Meta's table gives Luna scores on only four of six boards, with no index or cost-per-task figure. This article takes Luna first, then puts Ember back inside its own family, where its relationship with Spark and Wonder is a division of labour, not a rivalry.

Against GPT-6 Luna: four wins — but half the table is blank

Of the six component benchmarks (all Meta-reported, reproducing AA evaluations), Luna appears on four: AutomationBench-AA 37.8% vs 20.2%; OSWorld 2.1 69.5% vs 50.5% (the second-widest gap, 19 points); Agents' Last Exam 61.6 vs 57.1; PerceptionBench 80.6 vs 62.8 (a 17.8-point gap). Ember wins all four — and by wider margins than against Haiku. Terminal-Bench 4.0 and FrontierCode 2.0-AA have no Luna score in Meta's table. A blank means "Meta didn't list one", not that Luna scored zero — don't write it as "Luna scored zero on those two".

DimensionMuse EmberGPT-6 Luna
AutomationBench-AA37.8%20.2%
OSWorld 2.169.5%50.5%
Agents' Last Exam61.657.1
PerceptionBench80.662.8
Terminal-Bench 4.040.9%— (no score in Meta's table)
FrontierCode 2.0-AA27.2%— (no score in Meta's table)
AA Intelligence Index v5.036— (not confirmed in this round)
Cost per task (AA Index)$0.14— (not confirmed in this round)
List price (input / output)$0.10 / $0.50 (under 200K)$0.10 / $0.50

Mind the comparison's limits: Luna's AA Index and cost per task were not confirmed in this round and are left blank — this site won't invent numbers. So "Ember beats Luna on four boards" holds only at the component level; at the composite-index level the two have no comparable figures. On price, both list at $0.10/$0.50, but Ember has the 200K cliff; whether Luna has a similar tier was not confirmed in this round.

Inside the family: Ember trails Spark by one point, at a 17× price gap

AA Intelligence Index v5.0 (max effort): Spark 1.3 scores 37, Ember 36 — one point apart; cost per task: Spark $2.39, Ember $0.14 — a 17× gap. Alexandr Wang's own words: Ember delivers "nearly Muse Spark 1.3 (max) intelligence at a 17x cheaper price point (per AA)". That combination — one point, 17× — is the entire basis for Meta's two-model division of labour: the expensive model makes a few high-value decisions, the cheap one does the mass of repeatable execution. Simon Willison's hands-on test calls Ember 40× cheaper than Wonder — a different basis from Wang's 17× (per-task cost vs an unstated basis), so don't mix them.

DimensionMuse EmberMuse Spark 1.3Muse Wonder
AA Intelligence Index v5.0 (max)3637— (not confirmed in this round)
Cost per task (AA Index)$0.14$2.39— (not confirmed in this round)
Family positionSmallest (November)Mid-weight mainstay (April)Frontier flagship (October)
Open weightsNoNo (the open one is Glimmer, Spark's twin)No (Wonder Full stays locked inside Muse Code)
PositioningHigh-volume executionMid-weight mainstayPlanning and frontier tasks

The two-model split: how "Wonder plans, Ember executes" works in practice

Meta's positioning line is "Muse Wonder plans, Muse Ember handles the volume". In practice: Wonder / Spark plan, decompose and make high-value judgements, while Ember takes sub-agent tasks, tool calls, routing, extraction and long-document parsing — the high-volume, repeatable, low-value-per-call part. Muse Code is where the split lives: Ember is available on all plans, Wonder Full stays locked inside Muse Code, and Ember itself was trained with Wonder Full. The split only pays if the price gap is real: a 17× per-task cost gap makes "plan once, execute a thousand times" pencil out; if execution tasks output as much as planning tasks, the gap narrows and so does the benefit.

Three relationships not to confuse

Ember's family relationships are the easiest thing to get wrong; keep three straight. First, Ember vs Glimmer — Glimmer is the open-weight twin of Spark (August), Ember has no open weights, and the autocomplete suggestion "muse ember open weights" shows people asking, not weights existing. Second, Ember vs Spark — Ember is not a shrunken Spark but a different model with a different job in the same family: one index point apart, 17× apart on cost per task. Third, Ember vs Wonder Full — Ember was trained with Wonder Full, which itself is not offered externally and stays locked inside Muse Code.

FAQ

Is Ember stronger than GPT-6 Luna?

On the components, yes: on the four boards where Meta's table gives Luna a score (AutomationBench, OSWorld, Agents' Last Exam, PerceptionBench), Ember wins all four — OSWorld by 19 points, PerceptionBench by 17.8. But Luna's AA Index and cost per task were not confirmed in this round, so the two have no comparable composite figures; and Terminal-Bench and FrontierCode have no Luna score in Meta's table either.

Ember or Spark — which is stronger?

On the index, Spark 1.3 leads by one point (37 vs 36); on cost per task, Ember is 17× cheaper ($0.14 vs $2.39). The two are not rivals but a division of labour: Spark is the mid-weight mainstay, Ember the high-volume executor.

What does "Wonder plans, Ember executes" mean?

Meta's two-model pattern: Wonder / Spark plan, decompose and make high-value judgements; Ember takes sub-agent tasks, tool calls, routing, extraction and long-document parsing. Muse Code is where the split lives, with Ember on all plans.

Is Ember open source?

No. Ember ships no open weights and is API-only. Meta's open-weight Muse model is Glimmer / Spark Open (August) — a different model.

ModelBrief's take: The Ember-vs-Luna comparison looks one-sided but is actually incomplete — four wins are real, yet the index and cost-per-task rows are blank, and so are the Terminal-Bench and FrontierCode rows. The more informative comparison is inside the family: a one-point index gap against a 17× price gap defines Ember's place in the Muse lineup — not a weaker Spark, but the model that handles "volume" in the division of labour. To understand Ember, understand that split first, not just how many points it won against Luna.

Reference: Meta research blog — six component benchmarks, family positioning and the headline claim · Artificial Analysis — Index 36 / 37, cost per task $0.14 / $2.39 · Simon Willison hands-on test (40× cheaper than Wonder) (rewritten and organized by ModelBrief; facts per the AA evaluation and the Meta launch post)