Meta’s Superintelligence Lab announced Muse Ember on 9 November 2026. The company says it is its smallest and most efficient model yet, built for the routine work AI agents do all day – calling tools, handling coding subtasks, routing requests, extracting data and parsing long documents.
Muse Ember can read and generate text, images, audio and video, and it can process up to one million tokens in a single prompt. Meta makes it available now through the Meta Model API and Muse Code. Pricing is $0.10 per million input tokens and $0.50 per million output tokens, which translates to about $0.14 per task in the Artificial Analysis benchmark.
How it stacks up against competitors
In Meta’s own Artificial Analysis Intelligence Index, Ember scores 36 while costing $0.14 per task, far cheaper than the larger Muse Spark 1.3, which costs $2.39 for a near‑identical score of 37. Meta says Ember beats Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT‑6 Luna on cost‑per‑intelligence across most of its tests. However, on the FrontierCode 2.0‑AA coding benchmark, Haiku scores 34.9 percent versus Ember’s 27.2 percent, according to the company.
TechCrunch notes Ember’s price per million input tokens is comparable to open‑source models like DeepSeek V4.1 Flash and GLM 5.3 Flash, making it competitive at scale.
Where Ember fits in Meta’s model lineup
Ember is the fourth model in Meta’s Muse series. It follows the mid‑weight Spark (April 2026), its open‑source twin Glimmer (August 2026), and the frontier Wonder model (October 2026). Meta envisions Ember working alongside the larger Muse Wonder: the heavyweight model sets policies, while Ember applies them to tens of thousands of cases, such as in customer‑support automation.
Why it matters
For businesses that run large numbers of automated support or data‑extraction agents, Ember offers a cheaper way to run those repetitive steps without sacrificing the ability to work with text, images, audio or video. The lower cost per task could make high‑volume AI workflows more affordable, though its coding performance still trails some competitors.
