Meta Launches Muse Ember, Its Smallest and Cheapest Agent Model

What happened

The new model targets the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long-document parsing. Meta says Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting.

Ember launched the same day on Muse Code and the Meta Model API as muse-ember-1.0, with a 1M-token context window and up to 512K output tokens, enough to take in a full codebase or document set in one call. Meta says Ember is the first model trained using Muse Wonder Full.

Beyond pricing, Meta cited third-party scores: an Artificial Analysis Intelligence Index of 36, matching Claude Haiku 5.5, at about $0.14 per task; on the same leaderboard Muse Spark 1.3 scores 37 at $2.39 per task. The company says no model scores higher for less per task.

Key facts

1. Publisher and date: Meta released Muse Ember on Nov 9, the smallest and cheapest model in the Muse family
2. Pricing: $0.10 per million input tokens, $0.50 per million output tokens, $0.004 per million cached input tokens
3. Context and output: 1M-token context window, up to 512K output tokens per call
4. Third-party scores: Artificial Analysis Intelligence Index 36, matching Claude Haiku 5.5; about $0.14 per task versus $2.39 for Muse Spark 1.3
5. Vendor scores: Agents' Last Exam 61.6%, above Haiku 5.5's 57.0% and GPT-6 Luna's 57.1%; Terminal-Bench 4.0 40.9%, above Haiku's 39.2%; FrontierCode 2.0 27.2%, below Haiku's 34.9%
6. Training: The first model trained using Muse Wonder Full; Meta says pair Muse Wonder with hard problems and Ember with high-volume repeat work
7. Availability: Live the same day on Muse Code and the Meta Model API as muse-ember-1.0; also on Google Cloud and Oracle Cloud
8. Safety: Evaluated under Meta's Superintelligence Scaling Framework, chemical and biological, cybersecurity and loss-of-control risks all stayed below high-risk thresholds