muse ember: Meta's Small Model Built for Agent Execution

The muse ember release is Meta's answer to a cost problem every team running agents in production already knows: most agent tokens are not deep reasoning, they are the hundreds of small structured steps around it. On November 9, Meta introduced Muse Ember, its smallest and most efficient model, designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. Source

What Meta shipped

Muse Ember is available now on the Meta Model API at $0.10 per million input tokens and $0.50 per million output tokens. It carries a 1M-token context window and up to 512K output tokens, so it can take in a full codebase or document set in a single call. It is multimodal on input: text, image, audio, and video. Source

Meta positions it as the execution layer of a multi-model workflow. The model page describes pairing Muse Ember with Muse Wonder and Muse Spark, using the larger models for deep planning, reasoning, and orchestration while Ember handles high-volume execution. That is the subagent pattern given an official home: a big model decides what to do, a small model does the bulk of the doing. Source

The framing is a shift worth noting. Small models have always been cheaper per token; the argument Meta is making is that the executor tier is now a first-class product with its own benchmark sheet, not a fallback you tolerate. When a vendor publishes pass rates on end-to-end business workflows for its cheapest model, it is telling you where it expects the majority of agent tokens to go. Source

The cost curve Meta is claiming

Meta's headline claim is about intelligence per dollar: it says Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, matches Haiku's best score at about a third of the cost, and beats every Luna setting. The chart behind the claim uses the Artificial Analysis Intelligence Index against cost per task at each reasoning effort, with Ember's costs estimated at proposed pricing. Source

The published points make the claim concrete. On Meta's table, Muse Ember scores 33.9 at $0.116 per task and 35.8 at $0.136. Claude Haiku 5.5 ranges from 24.6 at $0.092 up to 36.2 at $0.411, and GPT-6 Luna from 21.4 at $0.073 to 27.8 at $0.152. So Ember's top setting lands within 0.4 points of Haiku's top setting at roughly a third of the per-task cost, while costing more per task than the cheapest Luna and Haiku settings that score far lower. These are Meta's own reported numbers from Artificial Analysis measurements, and Ember's per-task costs are estimates at list pricing, not metered bills. Source

The benchmark sheet

On agentic work, Meta reports Muse Ember (at max reasoning) scoring 37.8% on AutomationBench-AA end-to-end business workflows, against 13.9% for Claude Haiku 5.5 and 20.2% for GPT-6 Luna, a strict pass@1 measurement by Artificial Analysis. On Agents' Last Exam, Meta's internal evaluation on a Linux subset, Ember scores 61.6% against 57.0% for Haiku and 57.1% for Luna. On Meta's internal Personal Superintelligence Index, Ember leads 46.8% to 36.1% and 37.0%. Source

On coding, Ember scores 40.9% on Terminal-Bench 4.0 in its best published harness, against 39.2% for Haiku and 16.4% for Luna, and 46 on the Coding Agent Index-AA against 37 and 31. Haiku keeps one win on the sheet: FrontierCode 2.0-AA, real engineering tasks judged on mergeability, where Haiku's 34.9 beats Ember's 27.2. On the general Artificial Analysis Index, Ember and Haiku tie at 36, with Luna at 28. Mixed provenance matters here: some of these are third-party Artificial Analysis measurements, and some are Meta internal evaluations, as labeled in Meta's own post. Source

Multimodal and long documents

The multimodal results matter for document-heavy agent work. On OSWorld 2.1, a computer-use benchmark across desktop apps, Meta reports Ember at 69.5 partial and 33.8 binary, ahead of Haiku's 68.9 and 27.8 and well ahead of Luna's 50.5 and 17.0. On PerceptionBench with tools, a fine-grained visual perception test, Ember scores 80.6 against Haiku's 78.6 and Luna's 62.8. On GDP.pdf-AA, professional work over long documents, Ember and Luna tie at 22.8 with Haiku at 20.8. Source

Pair those numbers with the context window and the document-agent case writes itself: a 1M-token input with audio and video understanding, priced at $0.10 per million tokens, is built for the pass where an agent ingests a contract set, a meeting archive, or a repository before the expensive model reasons over the distilled result. Meta's own description of the workloads is blunt about this: routing, extraction, and long document parsing are listed first among the jobs agents repeat all day. Source

What early testers report

Meta published partner quotes alongside the launch. Cloudflare's CTO Dane Knecht says Muse Ember ran up to 30% faster at the tail than the comparable model they tested and handled broad, multi-repo investigation tasks well. You.com's inference lead reports a p95 latency of 2.2 seconds across coding quests, described as 4x to 13x faster than leading open-weight models, with 36 to 48% fewer output tokens. Cline's founder calls it purpose-built for subagent work like coding and document analysis, and Box's VP of AI Products describes it scoring contracts against a negotiation playbook clause by clause. These are vendor-selected early-access quotes, so treat them as directional, not independent reviews. Source

Our read

This section is our assessment, not Meta's claims. The strategic story is not that a small model beat Haiku on a chart; it is that the major labs now all sell the same architecture for agent systems: an expensive planner and a cheap executor. Anthropic has Haiku, OpenAI has Luna, Meta now has Ember, and the harnesses, from Cline to managed runtimes like the one we covered in mistral studio agents, are built to route subagent calls to the cheap tier. Model routing stops being a cost hack and becomes the default design.

Two practical notes. First, per-task cost is the right unit, and Meta is right to publish it, but the estimates come from list pricing on one benchmark harness; your routing decision should come from your own trace data, where the ratio of planner tokens to executor tokens is whatever your workload says it is. Second, a cheap high-volume executor is exactly the layer that touches the most untrusted content: tool results, web pages, documents. The injection risks we covered in AI agent memory poisoning apply at the subagent layer with more exposure, not less, because that is where untrusted text enters the system at scale. For Meta's own ecosystem, this is also the second Muse release with device implications we have tracked after the Muse Gadgets SDK; a 1M-context, low-cost execution model is the kind of component that eventually sits behind agent features on edge hardware, though Meta has announced no on-device version of Ember.

Bottom line: if you run agents at volume, Muse Ember is worth a benchmark slot against whatever Haiku-class model currently carries your subagent load. Run it on your own traces, at your own reasoning settings, before you route production traffic to it.