Many MindedMeta's Muse Ember matches a rival's benchmark score at a third of the cost
On November 9, 2026, Meta AI announced Muse Ember, which it calls its smallest and most efficient model. It is available as muse-ember-1.0 on Muse Code and the Meta Model API. Pricing is $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. Meta aims it at repeated agent work: tool calls, coding subtasks, routing, extraction and long-document parsing.
The specifications are large for a small model. It has a 1M-token context window, up to 512K output tokens, supports the Responses API, and natively handles text, images, audio and video. On the Artificial Analysis Intelligence Index, Ember at max scored 35.8 at $0.136 per task. Claude Haiku 5.5 at its top setting scored 36.2 at $0.411 per task, and GPT-6 Luna's best was 27.8 at $0.152. Against Meta's own Muse Spark 1.3, Ember scores 36 to 37 on that index, while cost per index task falls from $2.39 to $0.14.
The results are mixed. Ember led on AutomationBench-AA (37.8), Terminal-Bench 4.0 (40.9) and Coding Agent Index-AA (46). Haiku 5.5 led on FrontierCode 2.0-AA, 34.9 to Ember's 27.2.
Why it matters: agent workloads call a model many times, so price per task largely decides who can afford to run them. A model near the frontier of its size class at a third of the cost brings capable automation within reach of smaller builders and cheaper hardware budgets. That is access to intelligence getting cheaper.
Caveats: the cost comparisons are estimates at proposed or list pricing. Several results, including OSWorld 2.1 and PerceptionBench, are Meta internal evaluations. Terminal-Bench figures use each model's best published harness. Partner praise from Box, Cline, Oracle, Cloudflare and You.com comes from early testing and is vendor-supplied. Meta says its risk evaluations found Ember below high-risk thresholds; independent verification is worth watching for.
Source: Google News