Meta Launches Muse Ember, Its Smallest and Most Efficient Model: $0.10/$0.50 per Million Tokens, 1M Context, and an Artificial Analysis Score of 36 That Ties Claude Haiku 5.5 at About a Third of the Cost per Task

Background and Pain Points

A single agent task often breaks into dozens or hundreds of small steps: reading files, calling tools, running tests, extracting fields. Using a flagship model for every step is too slow and too expensive, so every lab now ships a cheap small model, such as Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna. Early on November 10 SGT, Meta Superintelligence Labs launched Muse Ember to compete directly in this tier. It completes Meta's lineup: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable coding and agent tasks at scale.

Architecture Highlights

Meta has not disclosed the parameter count, only that this is its smallest model so far and the first specialized model trained using Muse Wonder Full. It supports a 1M-token context window and up to 512K output tokens, so it can take in a whole codebase or document set in one call. It natively understands text, images, audio and video, so it can read screenshots, charts and scanned documents or work from voice memos and meeting recordings without a separate model. Meta says it received the same core safety training as larger models, and was tested on respecting user intent, resisting instructions hidden in web pages or tool outputs, and asking for human input before high-impact actions.

Benchmarks

Third-party results from Artificial Analysis: an Intelligence Index score of 36, tied with Claude Haiku 5.5 (max), at about $0.14 per task versus about $0.41 for Haiku; at xhigh effort it scores 34 for $0.12. Running in Muse Code, it scores 46 on the Coding Agent Index, ahead of Haiku 5.5 in Claude Code (37) and GPT-6 Luna in Codex (31). It reaches 40.9% on Terminal-Bench 4.0 (Haiku 5.5 39.2%), but only 27.2% on FrontierCode 2.0, behind Haiku's 34.9%. Knowledge is the main weakness: AA-Omniscience accuracy is 21.1% because it declines 64% of questions, which keeps its hallucination rate at 17.4%, about half of Haiku's. The Meta blog also lists OSWorld 2.1 (partial credit) at 69.5% and PerceptionBench at 80.6%; those are Meta internal evaluations.

Getting Started

It is available now on the Meta Model API as muse-ember-1.0, with Responses API support, and directly inside Muse Code. Pricing is $0.10 input, $0.004 cached input and $0.50 output per million tokens. Artificial Analysis notes that cached input makes up most of the cost in agentic evaluations, and Ember's cached price is 60% cheaper than similarly priced models. The natural use is as a subagent: plan with Muse Wonder or another flagship, then fan bounded subtasks out to Ember. It is not a good standalone choice for knowledge-heavy Q&A. The weights are not open, so it cannot be run locally.