Meta releases Muse Ember, a low-cost small AI model for agent work

Meta Superintelligence Labs released Muse Ember on Monday, November 9, its smallest and most cost-efficient model, built for the repetitive steps AI agents perform all day. It is available on the Meta Model API at $0.10 per million input tokens and $0.50 per million output tokens, and Meta claims it delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting.

The headline comparison is Meta’s own. It is built on Artificial Analysis results that Meta reproduces in its launch post, and some of Ember’s cost figures there are labelled as estimates. No independent hands-on testing beyond Meta’s selected launch partners is available yet.

What Muse Ember is built to do

Meta describes Ember as “designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing”. In plainer terms, it is meant to be the worker model inside an agent system, not the one that does the high-level planning.

Meta’s own framing is a two-tier setup: “Muse Wonder plans, Muse Ember handles the volume”. TechCrunch gives a customer-support example. A heavier model sets the rules for handling a type of issue, and a lighter model then applies those rules across tens of thousands of cases. Meta says agents spend most of their tokens on many small, structured steps. It says Ember can carry a business workflow end to end, pulling records, calling tools in sequence and updating systems across apps.

Ember also handles more than text. According to TechCrunch, Meta trained it to work across text, images and audio. In Meta’s words, “agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model”. For builders, that could remove a separate transcription or vision step from some pipelines, if the quality holds up in practice.

How Muse Ember compares with Haiku 5.5 and GPT-6 Luna

Meta’s central claim concerns cost per task, meaning the total spend to complete a benchmark task, not the price per token. Meta says Ember “matches Haiku’s best score at about a third of the cost and beats every Luna setting,” and that “no model scores higher for less per task” on the full Artificial Analysis leaderboard.

The chart data in Meta’s post shows this:

• Muse Ember: 33.9 at $0.116 per task, and 35.8 at $0.136 per task.
• Claude Haiku 5.5: ranges from 24.6 at $0.092 up to 36.2 at $0.411 per task.
• GPT-6 Luna: ranges from 21.4 at $0.073 up to 27.8 at $0.152 per task.

On those figures, Ember’s top point (35.8) sits just below Haiku’s top point (36.2), at roughly a third of the cost per task. The chart notes say Ember’s costs are “estimates at proposed pricing.” The same chart’s footnote says “Muse Ember cost at list pricing”. Both labels appear on the same chart, and the post does not explain which applies.

Meta also published a table of results at each model’s maximum reasoning setting. “AA” in a benchmark name marks an Artificial Analysis version:

Benchmark (max reasoning)	Muse Ember	Claude Haiku 5.5	GPT-6 Luna
AutomationBench-AA (business workflows)	37.8	13.9	20.2
Agents’ Last Exam (end-to-end agent tasks)	61.6	57.0	57.1
Terminal-Bench 4.0 (agentic terminal coding)	40.9	39.2	16.4
Coding Agent Index-AA	46	37	31
FrontierCode 2.0-AA (mergeable engineering tasks)	27.2	34.9	25.2
GDP.pdf-AA (long-document work)	22.8	20.8	22.8
OSWorld 2.1, partial / binary (desktop computer use)	69.5 / 33.8	68.9 / 27.8	50.5 / 17.0
PerceptionBench with tools (visual perception)	80.6	78.6	62.8
Artificial Analysis Index	36	36	28

All figures in the table are from Meta’s announcement.

Ember does not lead everywhere in Meta’s table. Haiku 5.5 scores higher on FrontierCode 2.0-AA, a test judged on whether code would be mergeable. Ember ties Luna on GDP.pdf-AA and ties Haiku on the overall Artificial Analysis Index. Meta also lists a “Personal Superintelligence Index,” which it describes as an internal composite of personal and knowledge work, where Ember scores 46.8 against 36.1 and 37.0. Because that index is internal, outside parties cannot reproduce it.

TechCrunch adds that, at 10 cents per million input tokens, Ember is cost-competitive with open-source rivals such as DeepSeek V4.1 Flash and GLM 5.3 Flash. No benchmark scores for those models appear in the sources.

Price, context window and access

Spec	Muse Ember
Input price	$0.10 per million tokens
Output price	$0.50 per million tokens
Context window	1M tokens
Maximum output	Up to 512K tokens
Inputs	Text, images, audio
Availability	Meta Model API, from November 9

Meta says the 1M-token window lets Ember take in a full codebase or document set in a single call. The announcement names the Meta Model API as the place to use Ember. It does not describe open weights for this model.

Developers new to that API should check its current terms. As of August 10, a WaveSpeedAI explainer described the Meta Model API as a public preview framed for US developers, with self-serve access and free credits. The same piece notes that Meta says developers can point existing OpenAI SDK code at the API with a new key. It cautions that compatible APIs can still differ in error handling, tool-call behaviour and rate-limit headers. The sources do not say whether the preview status or regional limits have changed since August.

What early partners report

Meta’s post includes statements from five early testers: Box, Cline, Oracle, Cloudflare and You.com. These are partner quotes the maker selected, not independent evaluations, but some include specific numbers:

• Cloudflare: CTO Dane Knecht says Ember “ran up to 30% faster at the tail than the comparable model we tested.” “The tail” refers to the slowest responses. He also says Ember handled broad, multi-repo investigation tasks well. Cloudflare does not name the comparison model.
• You.com: Inference lead Chak Pothina says Ember completed every coding quest in its testing with a p95 latency of 2.2 seconds. p95 is the time within which 95% of responses arrive. He calls this “about 4x to 13x faster than leading open-weight models, and with 36–48% fewer output tokens”.
• Box: VP of AI Products Yashodha Bhavnani says Ember can score a contract against a negotiation playbook clause by clause and arrive at the right overall risk rating, on the Box AI Complex Work Eval.
• Cline: Founder Saoud Rizwan calls it “purpose-built for, and excels at, subagent work like coding and document analysis”.
• Oracle: SVP Konstantinos Papamiltiadis describes “an encouraging ability to tackle complex engineering tasks” in early testing.

Where Ember fits in the Muse lineup

Ember is the fourth model under Meta’s Muse brand. The earlier three are:

• the mid-weight Spark, released in April;
• Glimmer, an open-source twin of Spark, released in August;
• Wonder, the frontier model, released in October.

Meta’s consumer product, the Muse Agent, launched in September.

The price gap with Wonder is large. Virtual Arc reports that the API version of Muse Wonder costs $4 per million input tokens and $20 per million output tokens. It also reports that the stronger Muse Wonder Full is restricted to Muse Code subscriptions. That makes Ember one-fortieth of Wonder’s listed per-token price on both input and output, which fits Meta’s plan-then-execute pairing.

Ember is not the only cheap Muse option. A ToolScout.ai catalog snapshot dated October 31 lists Muse Spark 1.3 and Spark 1.2 at $0.10 per million input tokens. That is the same input price as Ember. Meta’s announcement does not compare Ember with Spark, so builders already on Spark have no official guidance on whether to switch.

What it means for builders, and the open questions

For teams running agents at volume, Ember targets the part of the bill that grows fastest: the many small tool calls, extractions and routing decisions that sit beneath a planning model. If Meta’s cost-per-task figures hold up on real workloads, a two-tier stack with a frontier planner and Ember as the subagent could cut spend without a large drop in task success.

Several caveats apply:

• The evidence is the maker’s. The comparison charts come from Meta’s announcement. Some Ember costs are labelled as estimates, and one index is internal to Meta.
• Per-token price is not per-task cost. WaveSpeedAI’s guidance on the Meta Model API warns that a low token price can still become expensive with long context, tool loops, retries and verbose outputs. Ember’s 512K output ceiling and reasoning settings make task-level measurement important.
• Coding results are mixed. Ember leads on Terminal-Bench 4.0 and Coding Agent Index-AA but trails Haiku 5.5 on FrontierCode 2.0-AA in Meta’s own table. Teams using it for code that must merge cleanly should test on their own repositories.
• API maturity. Earlier reporting described the Meta Model API as a US-focused public preview. Production teams should confirm regions, quotas and data terms before moving customer traffic.

What to watch next

The next signals are independent ones. Watch for Artificial Analysis publishing Ember’s results on its own leaderboard, and for third-party tests of the cost-per-task claim at final pricing. Also watch for hands-on reports on Ember’s audio and video input, which Meta highlights but does not benchmark separately in its table. Two open questions remain: whether Meta brings Ember to other platforms beyond its own API, and whether it pairs Ember with an open-weight sibling the way Glimmer paired with Spark.