Another tough contender has squeezed into the lightweight model race. On November 10, Meta launched Muse Ember, billing it as its "smallest and most efficient" model yet, built for complex reasoning and everyday tasks that agents execute over and over. It does not compete on parameter count—only on price-performance.
Matches Haiku on benchmarks, at a fraction of the bill
On the Artificial Analysis leaderboard, Muse Ember's overall performance can already go toe-to-toe with Claude Haiku 5.5, yet its per-task cost is only one-third of its rival's. It also beats GPT-6 Luna in multiple tests. For example, on the end-to-end business workflow AutomationBench-AA, it scored 37.8, far ahead of Haiku's 13.9 and Luna's 20.2. It also leads in areas such as agentic coding and long-document processing.
Swallows an entire codebase in one go
Ember comes with a 1 million-token context window and can output up to 512,000 tokens, meaning a single call can ingest an entire codebase or a whole stack of documents. Agents can carry out hundreds of coding subtasks in a single session, reading files, running tests, and making targeted bug fixes in one smooth flow.
It is also a native all-rounder: it handles text, images, audio, and video alike, and can take on screenshots, charts, scans, meeting recordings, and videos—no need to bolt on another model. On safety, it was evaluated under the Meta SI framework before release, and its risk is below the high-risk red line. The pricing is affordable: $0.10 per million input tokens and $0.50 per million output tokens, and it is already available on the Meta Model API, Google Cloud, and Oracle Cloud.
