Meta Launches Muse Ember — Lightweight Agent Model Trained by Flagship Muse Wonder, Matches Haiku 5.5 at ⅓ Cost
Meta Launches Muse Ember — Lightweight Agent Model Trained by Flagship Muse Wonder
Meta released Muse Ember on November 9, 2026 — its smallest, most efficient model designed for high-volume agentic subtasks: tool calls, coding subtasks, routing, extraction, and long document parsing. At $0.10/$0.50 per M tokens (input/output), it matches Claude Haiku 5.5 at ~⅓ the cost and beats GPT-6 Luna on Artificial Analysis benchmarks.
Breakthrough: Muse Ember is the first model Meta trained using Muse Wonder Full (its flagship frontier model) — a concrete step toward recursive self-improvement where frontier models automate the training of specialized downstream models.
Model Portfolio: Wonder → Spark → Ember
| Model | Role | Context | Input/Output (per M) | Artificial Analysis Score | Per-Task Cost |
|---|---|---|---|---|---|
| Muse Wonder Full | Planner / Hardest problems | 1M+ | Not public | Frontier (~59) | Premium |
| Muse Wonder | Hard problems | 1M | API pricing | High | High |
| Muse Spark 1.3 | Coding / agentic workflows | 1M | Mid-tier | Strong coding | Medium |
| Muse Glimmer | Open-weight twin of Spark | 1M | Free (weights) | Similar to Spark | Free |
| Muse Ember | Volume agent subtasks | 1M | $0.10 / $0.50 | 36 (ties Haiku 5.5 max) | ~$0.14/task (⅓ Haiku) |
Meta's framing: "Muse Wonder plans, Muse Ember handles the volume." — A two-tier agentic architecture where the flagship models strategy/planning and the lightweight model executes repetitive structured steps.
Benchmark Performance: Efficiency Frontier
| Benchmark | Muse Ember | Claude Haiku 5.5 (max) | GPT-6 Luna (max) | DeepSeek V4.1 Flash |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 36 | 36 | 28 | ~34 |
| AutomationBench-AA | 37.8 | 13.9 | 20.2 | — |
| FrontierCode 2.0-AA | 27.2 | 34.9 | — | — |
| Terminal-Bench (coding) | Competitive | Strong | Baseline | Strong |
| Multimodal (text/image/audio/video) | Native | Limited | Limited | Text-only |
| Context window | 1M tokens | 200K | 128K | 128K |
| Max output tokens | 512K | 128K | 128K | 128K |
| Estimated cost per AA task | $0.136 | $0.411 | ~$0.35 | ~$0.15 |
Key wins: 2.9× cheaper than Haiku, 4× context/output vs rivals, native multimodal (text/image/audio/video), best-in-class AutomationBench.
Gap: FrontierCode 2.0 (real-repo coding) trails Haiku 5.5 — not a drop-in for complex mergeable code changes.
Recursive Training: "Trained by Muse Wonder Full"
| Aspect | Detail |
|---|---|
| Claim | First model trained using Muse Wonder Full (flagship) |
| Possible automation | Post-training (RLHF, DPO, preference learning), data curation, eval generation |
| Thinking Machines Labs precedent | Demonstrated small LLM can run post-training end-to-end (Aug 2026) |
| Anthropic parallel | Claude does 26% of work building next model (AIPress Post 24) |
| Meta's disclosure | "Did not specify which part of workflow automated" |
Significance: If Wonder Full automated post-training/data curation for Ember, this is recursive self-improvement in production — the frontier model building its own specialized descendants. Meta has done this before (Wonder → specialized models for Muse Agent), but Ember is the first API-released model from this pipeline.
Multimodal Native: One Model for All Inputs
Muse Ember natively accepts text, images, audio, and video — no separate encoder/processor needed:
- Agents can read: screenshots, charts, scanned documents, PDFs
- Agents can process: voice memos, meeting recordings, video
- Desktop control: Operates apps from screenshots (computer-use capability)
- Single call: 1M-token context fits full codebase + docs + media
vs. competitors: Haiku 5.5 / Luna require separate vision/encoders; Ember = unified.
Availability & Pricing: API-Only, Hosted
| Platform | Model ID | Pricing |
|---|---|---|
| Meta Model API | muse-ember-1.0 |
$0.10 in / $0.50 out / $0.004 cached |
| Muse Code | Integrated | Same |
| Google Cloud | Vertex AI | Same |
| Oracle Cloud | OCI | Same |
| OpenRouter / neoclouds | Expected soon | Similar |
Closed weights: No local download, no quantization, no on-prem — hosted only. This contrasts with Glimmer (open weights) and DeepSeek/Gemma (open).
JiTTests: Dogfooding LLM-Generated Testing
Meta's JiTTests system (LLM-generated testing) caught 74 privacy SEVs in H1 2026 with 60% true-positive rate — evidence that Meta's own models are stress-testing each other in production. Ember benefits from this automated quality pipeline.
Competitive Landscape: The "Small Model for Agents" Race
| Model | Provider | Input/Output | Context | Key Feature | Status |
|---|---|---|---|---|---|
| Muse Ember | Meta | $0.10/$0.50 | 1M / 512K | Native multimodal, Wonder-trained | Live Nov 9 |
| Claude Haiku 5.5 | Anthropic | ~$0.25/$1.25 | 200K | Strong coding, Verification Program | Live |
| GPT-6 Luna | OpenAI | ~$0.15/$0.75 | 128K | Dots agent integration, Wallet | Live |
| Gemini 4 Hydrogen | $1.50/$7.50 | 1M | Free tier for Search, 57% Terminal-Bench | Live Nov 12 | |
| DeepSeek V4.1 Flash | DeepSeek | ~$0.10/$0.40 | 128K | Open weights, beats own flagship | Live |
| GLM 5.3 Flash | Zhipu | Competitive | 128K | Chinese market focus | Live |
Ember's niche: Lowest cost-per-task for Haiku-level intelligence + native multimodal + Wonder-trained pedigree. Best for volume agent fleets where per-task cost compounds.
Strategic Implications
- Recursive training proven at scale — Wonder Full → Ember pipeline works; expect more specialized models
- Two-tier agentic architecture — Planner (Wonder) + Executor (Ember) becomes Meta's default pattern
- Closed weights = recurring revenue — API-only model locks in hosting margin vs. open-weight commoditization
- Multimodal unification — Single model for text/image/audio/video reduces agent stack complexity
- Cost curve bending — $0.14/task makes agent fleets economically viable for SMB, not just enterprise
What Comes Next
- More Wonder-trained models — Specialized variants for legal, medical, financial, creative domains
- Ember on neoclouds — OpenRouter, Together, Fireworks, Lambda deployment
- Muse Agent integration — Ember as default executor for Meta's consumer agent (Spaces, video chat)
- Open-weight Ember? — Glimmer precedent suggests possible future release
- Benchmark response — Anthropic/OpenAI/Google likely to cut Haiku/Luna/Hydrogen pricing
- Recursive training transparency — Industry pressure to disclose what % of pipeline is automated
Related AIPress Coverage
- Meta Is Building an In-House AI Chip
- Meta Patches Muse Exploit That Let Attackers Control the AI Agent
- Recursive Self-Improvement Is No Longer a Thought Experiment
- Anthropic Launches Fable 5.2 and Mythos 5.2
Sources
- Meta Research Blog (Nov 9, 2026): "Introducing Our Smallest, Most Efficient Model: Muse Ember"
- The New Stack (Nov 9, 2026): "Meta launches Muse Ember to take on Claude Haiku and GPT-6 Luna"
- TechCrunch (Nov 9, 2026): "Meta releases Muse Ember, a new lightweight model for agent execution"
- SiliconANGLE (Nov 9, 2026): "Meta introduces Muse Ember to power cost-efficient AI agents"
- VentureBeat (Nov 9, 2026): "Meta's proprietary Muse Ember nearly matches Claude Haiku 5.5 at 1/3 the estimated per-task cost"
- MarkTechPost (Nov 9, 2026): Detailed benchmark analysis
- LinkedIn / Meta AI (Nov 9, 2026): Team announcements
- Artificial Analysis Intelligence Index (Nov 2026): Leaderboard data