Meta launched Muse Ember on 9 November, the smallest and cheapest model in its Muse family, built for the repetitive work AI agents do all day: tool calls, coding subtasks, routing, extraction and long-document parsing. It is live on the Meta Model API and in Muse Code as muse-ember-1.0, priced at $0.10 per million input tokens and $0.50 per million output tokens.
The pitch is intelligence per dollar. On Artificial Analysis's Intelligence Index v5, Ember scores 36 at maximum reasoning โ effectively tied with Claude Haiku 5.5 (36.2) and ahead of OpenAI's GPT-6 Luna (28) โ at about $0.14 per task against Haiku's $0.41. Meta says no model on the full leaderboard scores higher for less per task. Those are benchmark-provider figures, not independent hands-on results.
Ember takes text, images, audio and video, with a 1 million-token context window and up to 512,000 output tokens, and supports the Responses API. Meta does not disclose a parameter count and has not released the weights โ this is an API-only model, not an open download.
Meta frames Ember as the volume worker under its largest model, Muse Wonder: โMuse Wonder plans, Muse Ember handles the volume.โ It is also the first model Meta says was trained using Muse Wonder Full, its newest training approach. Early partners named by Meta include Box, Cline, Oracle, Cloudflare and You.com.
The launch matters because agent workloads are mostly repetition โ and repetition is where per-token price decides whether an agent product can afford to run. At $0.10 per million input tokens, Ember is priced to be called thousands of times a day; whether it holds up in real agent loops, rather than on launch-day benchmarks, is the test that starts now.