Meta introduced Muse Ember on Monday, the smallest and most efficient model in its Muse family, aimed squarely at the routine work that AI agents perform all day: tool calls, coding subtasks, routing, extraction, and parsing long documents. The model went live on the Meta Model API at $0.10 per million input tokens and $0.50 per million output tokens.

According to Meta's announcement, Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting. The company says the model matches Haiku 5.5's best score at roughly a third of the cost and beats every Luna setting, placing it on the efficiency frontier of the Artificial Analysis leaderboard, where no model scores higher for less money per task. For more context on this story, see our ongoing AI trends.

A Model Built for Agent Subtasks

Meta's pitch is that modern agents spend the bulk of their tokens on many small, structured steps rather than one giant answer, and that Muse Ember is engineered for exactly that workload. The company describes a model that can carry a business workflow end to end, pulling records, calling tools in sequence, and updating systems across applications.

The context window is unusually generous for this class of model. Muse Ember accepts up to 1 million input tokens and can produce up to 512,000 output tokens in a single call, enough to ingest a full codebase or a large document set at once.

Benchmarks: Strong Scores, One Notable Gap

The headline numbers come from Artificial Analysis data dated November 9, 2026, reproduced in Meta's announcement. On the Artificial Analysis Intelligence Index, Muse Ember scored 33.9 at $0.116 per task and 35.8 at $0.136 per task across its highest reasoning settings. Claude Haiku 5.5's best result was 36.2, but it cost $0.411 per task, roughly three times as much. GPT-6 Luna topped out at 27.8. On the overall index at max reasoning, Muse Ember and Haiku 5.5 tied at 36, with Luna at 28.

Task-level results show the same pattern:

  • On Terminal-Bench 4.0, which measures agentic terminal coding, Muse Ember scored 40.9 at max reasoning, ahead of Claude Haiku 5.5 at 39.2 and far clear of GPT-6 Luna at 16.4.
  • On AutomationBench-AA, an end-to-end business workflow benchmark scored with strict pass@1, Muse Ember reached 37.8 versus 20.2 for GPT-6 Luna and 13.9 for Claude Haiku 5.5.
  • On Agents' Last Exam, a Meta internal evaluation of end-to-end agent tasks on the Linux subset, Muse Ember scored 61.6, ahead of Luna at 57.1 and Haiku 5.5 at 57.0.
  • On Meta's Personal Superintelligence Index, an internal composite of personal and knowledge work, Muse Ember scored 46.8 versus 37.0 for Luna and 36.1 for Haiku 5.5.
  • On OSWorld 2.1, a computer-use benchmark across desktop applications, Muse Ember scored 69.5 in partial-match scoring and 27.8 in binary scoring.
  • On PerceptionBench with tools, a fine-grained visual perception test, Muse Ember led with 80.6 against 78.6 for Haiku 5.5 and 62.8 for Luna.

There is one gap worth flagging. On FrontierCode 2.0-AA, which judges real engineering tasks on whether the resulting code can merge, Claude Haiku 5.5 led at 34.9, with Muse Ember at 27.2 and GPT-6 Luna at 25.2. It is a reminder that benchmarks diverge and that teams should match models to their actual workloads rather than a single leaderboard.

What the Efficiency Numbers Mean

The cost-per-task framing is where Meta is drawing the sharpest contrast. Haiku 5.5 needed to spend up to $0.411 per task at its highest reasoning setting to reach its best score of 36.2, while Muse Ember posted 35.8 at $0.136. Luna's best score of 27.8 came at $0.152 per task, above Muse Ember's price and well below its scores. For agent harnesses that run hundreds of subtasks per session, per-task cost compounds quickly, which is exactly the segment Meta is targeting.

What Early Partners Report

Meta's announcement includes early results from five partners: Box, Cline, Oracle, Cloudflare, and You.com.

Cloudflare CTO Dane Knecht said Muse Ember ran "up to 30% faster at the tail than the comparable model we tested" and handled broad, multi-repo investigation tasks well, calling it a strong fit for high-throughput agentic workloads.

At You.com, inference lead Chak Pothina reported that Muse Ember completed every coding quest in the company's testing with the lowest latency of any model tried, a p95 of 2.2 seconds. That is, according to Pothina, about 4x to 13x faster than leading open-weight models, with 36% to 48% fewer output tokens.

Box VP of AI Products Yashodha Bhavnani said that on Box's AI Complex Work Eval, the model could score a contract against a company's negotiation playbook clause by clause and land on the correct overall risk rating. Cline founder Saoud Rizwan called Muse Ember "purpose-built" for subagent work like coding and document analysis, and Oracle SVP Konstantinos Papamiltiadis cited strong potential for engineering teams.

Notable Points About Availability and Weights

Muse Ember is available through the Meta Model API as of Monday. The announcement does not mention open-weight availability, so teams that self-host should treat this as an API-only release for now. Meta Superintelligence Labs is credited with the work in partner quotes.

For developers assembling agent harnesses, the arrival of a cheap, fast subagent model adds a third major option alongside Claude Haiku 5.5 and GPT-6 Luna, both of which Muse Ember now undercuts on price at comparable quality on the benchmarks Meta cites. Whether that advantage holds under independent testing, outside data Meta reproduced from Artificial Analysis, will be the next thing to watch.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →