Meta Releases Muse Ember, a Lightweight Model Built for Agent Execution at Scale

Meta's Superintelligence Lab released Muse Ember on November 9, calling it the company's smallest and most efficient model to date. The model is designed explicitly for the high-volume, repetitive work that AI agents perform: tool calls, coding subtasks, routing, extraction, and long-document parsing [1]. It is the fourth model in the Muse family, following Spark (April), Glimmer (August), and the frontier model Wonder (October) [2].

Pricing and the Execution-Layer Pitch

Muse Ember is priced at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens [3]. Meta positions it as the execution layer in a two-model agent stack: "Muse Wonder plans, Muse Ember handles the volume" [6]. In this pattern, the expensive frontier model (Wonder, at $4/$20 per million input/output) decides what to do once; the cheap model carries out the resulting steps thousands of times.

The cached-input rate is the standout figure. At $0.004 per million tokens, it is one twenty-fifth the fresh-input price. Agent frameworks typically resend the same system prompt, tool schemas, and reference material on every call; caching that prefix can cut the marginal cost of a call dramatically [3]. However, the announcement does not specify cache write fees or expiry rules, so the effective savings depend on implementation details not yet public.

There is also a price tier change at 200,000 tokens: prompts above that length are billed at $0.30 input, $0.012 cached, and $1.50 output — three times the headline rate [5]. Meta's launch post quotes only the lower rate; the higher tier appears in the developer pricing table.

Benchmarks: What Meta Shows and What It Doesn't

Meta published a suite of benchmark results comparing Ember (at max reasoning) against Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna. The results are a mix of independent Artificial Analysis runs and Meta's own evaluations [4][6].

On the Artificial Analysis Intelligence Index, Ember and Haiku both score 36 (Haiku slightly higher before rounding), while Luna scores 28 [6]. Artificial Analysis estimates Ember's cost per task at $0.14 versus Haiku's $0.41 and Luna's $0.15, driven largely by output token volume: Ember produced 115 million output tokens across the index run, Haiku 298 million, at the same $0.50 output rate [5].

Ember leads on AutomationBench-AA (business workflows: 37.8% vs. 13.9% for Haiku), Coding Agent Index-AA (46 vs. 37), Terminal-Bench 4.0 (40.9% vs. 39.2%), OSWorld 2.1 computer use (69.5% partial credit vs. 68.9%), and PerceptionBench with tools (80.6% vs. 78.6%) [6]. It ties Luna on GDP.pdf long-document parsing (22.8% each), while Haiku scores 20.8%.

Haiku beats Ember on FrontierCode 2.0-AA, the benchmark that grades real engineering tasks on mergeability: 34.9% vs. 27.2% [6]. That test is the closest to a production pull-request evaluation. Meta's evaluation report notes that the Terminal-Bench row mixes sources: Ember's score comes from an Artificial Analysis run in the mini-swe-agent harness, while Haiku's is Anthropic's self-reported Claude Code result [5]. Four of the ten published rows are Meta's own runs, which the report describes as "best-effort" for competitors and warns that prompts and tooling "may not be optimally tuned for proprietary models" [5].

Multimodal and Availability

Ember natively processes text, images, audio, and video, allowing agents to work from screenshots, charts, scanned documents, voice memos, meeting recordings, and video without a separate model [6]. The Artificial Analysis model page lists only text and image inputs, so audio and video support rests on Meta's claim [5].

The model is available today via the Meta Model API and in Muse Code (Meta's terminal coding agent) under the ID muse-ember-1.0 [3]. It is not in the consumer Meta AI app or the Muse Agent product. Context window is 1 million tokens; maximum output is 512,000 tokens [3].

Safety and Partner Quotes

Meta says Ember was evaluated under its Superintelligence Scaling Framework for chemical/biological, cybersecurity, and loss-of-control risks, and remained below high-risk thresholds [6]. Agent-behavior evaluations covered respect for user intent, sensitive-information handling, resistance to hidden instructions, and seeking human input before high-impact actions.

The announcement includes five partner quotations [6]. Secondary coverage identifies two of them as coming from Cloudflare (CTO Dane Knecht) and You.com (inference lead Chak Pothina) [4][5]. The Cloudflare and You.com speaker names are provided by those secondary reports, not by the announcement's quote blocks themselves. The first quote in the announcement references the "Box AI Complex Work Eval" but does not identify the speaker. Two further quotations in the announcement carry no company or speaker identification in any available source.

Cloudflare reports Ember ran "up to 30% faster at the tail" and handled multi-repo investigations well; You.com measured 95th-percentile latency of 2.2 seconds on coding tasks with 36–48% fewer output tokens than other models tested [4][5].

Market Context

McKinsey's 2025 State of AI survey found 23% of organizations scaling agentic systems and 39% experimenting — roughly 62% combined [3]. In tech, 24% reported using agents in software engineering and 22% in IT. That existing usage is the addressable market for a model priced to make high-volume agent workloads viable.

What Remains Unknown

• No independent third-party has yet replicated Meta's internal benchmark runs.
• Cache write fees, TTL, and regional availability for the Meta Model API are not detailed in the launch materials.
• Long-context recall quality at the full 1 million tokens is untested publicly.
• The "trained using Muse Wonder Full" claim is not independently verified; it is Meta's description of its distillation approach [6].

For developers, the practical next step is to test Ember on a high-volume, easy-to-score task — routing or field extraction — against the current model, measuring accuracy, latency, and cost per thousand calls including retries. The bill, not the sticker price, is what the execution-layer argument rests on.