A one-point gap and a seventeenfold price gap
Meta released Muse Ember on 9 November, calling it its smallest and most efficient model, built for “the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing”. The comparison the company leads with is not against a rival but against itself. In its announcement, Meta puts Muse Ember at 36 on the Artificial Analysis Intelligence Index against 37 for its own Muse Spark 1.3 — and at about $0.14 per task against $2.39.
That is roughly a seventeenfold difference in cost for one index point. The token prices behind it are $0.10 per million input tokens and $0.50 per million output for Muse Ember, against $1.25 and $4.25 for Muse Spark 1.3. Cached input is $0.004 per million. Prompts over 200,000 tokens are billed differently, at $0.30 input, $0.012 cached input and $1.50 output per million.

Where the numbers come from
The index figure is independent: Artificial Analysis published Intelligence Index v5 on 9 November, and the cost-per-task comparison uses each model’s list pricing. On that index Muse Ember ties Claude Haiku 5.5 at 36, with GPT-6 Luna at 28. Meta’s own chart of intelligence against cost per task, drawn from the same source, shows Muse Ember at 35.8 for $0.136 a task at max reasoning, while Haiku 5.5 reaches 36.2 at $0.411.
The benchmark table is a mix. Scores carrying the “-AA” suffix come from Artificial Analysis: Muse Ember takes AutomationBench-AA at 37.8% against Haiku 5.5’s 13.9% and GPT-6 Luna’s 20.2%, and the Coding Agent Index-AA at 46 against 37 and 31 — but loses FrontierCode 2.0-AA, where Haiku 5.5 scores 34.9% to Muse Ember’s 27.2%. Agents’ Last Exam (61.6%), the Personal Superintelligence Index (46.8%) and PerceptionBench (80.6%) are labelled Meta internal evaluations. On Terminal-Bench 4.0 Meta reports 40.9% against 39.2% for Haiku 5.5 and 16.4% for GPT-6 Luna, with each model run in its own best published harness — mini-swe-agent, Claude Code and Codex respectively — which is a fair method and also one that makes the numbers harder to compare directly.
Built from the big model
Muse Ember is the first model Meta has trained using Muse Wonder Full, the frontier model it introduced in October, and the company frames it as delivery on a promise made then: to use the large models to produce specialised ones it could put everywhere. The line-up it describes now is Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable work at scale.

The context window is one million tokens with up to 512,000 output tokens, enough, Meta says, to take a full codebase or document set in a single call. The model is available on Muse Code and the Meta Model API as muse-ember-1.0, supports the Responses API, and is listed on Google Cloud and Oracle Cloud alongside Muse Spark and Muse Wonder.
Safety, and the partner numbers
Meta said it evaluated Muse Ember under its Meta Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks, and that across the categories assessed the model remained below the framework’s high-risk thresholds. It also says it tested how the model behaves acting on a user’s behalf, including whether it resists instructions hidden in web pages, documents or tool outputs. No scores for any of this are published.
Figures from early partners are their own. Cloudflare’s chief technology officer Dane Knecht said Muse Ember “ran up to 30% faster at the tail than the comparable model we tested”. You.com’s inference lead Chak Pothina reported a p95 latency of 2.2 seconds, “about 4x to 13x faster than leading open-weight models, and with 36–48% fewer output tokens”. Neither claim names the models compared against.
What to watch is whether the index position holds as Artificial Analysis reruns v5 on other harnesses. A tie with Haiku 5.5 at a third of the cost is the whole argument for this model, and it rests on one number from one leaderboard published on launch day.