Meta Releases Muse Ember, Its Smallest and Cheapest Model Yet
Muse Ember is built for the repetitive work AI agents do all day, matching Claude Haiku 5.5 on Artificial Analysis's index at about a third of the cost per task.
By Icarus3 min read
Meta released Muse Ember today, the smallest model in the Muse family and the cheapest way to run the repetitive work AI agents do all day. The model is available now through the Meta Model API and Muse Code as muse-ember-1.0, and it is also listed on OpenRouter.
Ember is built for tool calls, coding subtasks, routing, extraction, and long document parsing, the steps an agent repeats hundreds of times in a session. It is the first model trained using Muse Wonder Full, the larger frontier model introduced in October. It rounds out the Muse lineup: Wonder handles the hardest problems, Spark handles coding and agentic workflows, and Ember handles repeatable tasks at scale.
The Price Is the Point
Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. On Artificial Analysis’s Intelligence Index, it scores 36 at its maximum reasoning setting, matching Claude Haiku 5.5’s best score at about a third of the cost per task: $0.14 against $0.41. It beats GPT-6 Luna, which scores 28, at every reasoning setting, and Artificial Analysis places Ember on the efficiency frontier of its full leaderboard, meaning no model scores higher for less per task.
The gap with Muse Spark 1.3 is narrow on capability and wide on cost. Spark 1.3 scores 37 on the same index, one point above Ember, but costs $2.39 per task against Ember’s $0.14, with list prices of $1.25 per million input tokens and $4.25 per million output tokens.
Ember carries a 1M-token context window and can produce up to 512K output tokens, enough to take in a full codebase or document set in one call. It natively understands text, images, audio, and video, and it supports the Responses API.
Where Ember Wins, and Where It Doesn’t
On AutomationBench-AA, which measures end-to-end business workflows, Ember scores 37.8, against 13.9 for Haiku 5.5 and 20.2 for Luna, using strict pass@1 scoring. On the Coding Agent Index, which measures coding inside real agent harnesses, Ember scores 46, ahead of Haiku’s 37 and Luna’s 31. On Terminal-Bench 4.0, Ember scores 40.9, ahead of Haiku’s 39.2 and Luna’s 16.4.
Ember trails on other tests. On FrontierCode 2.0, which judges real engineering work on whether it could be merged, Ember scores 27.2, behind Haiku’s 34.9. And Artificial Analysis flags knowledge as the model’s main weakness: Ember’s accuracy on its AA-Omniscience test is 21.1%, the lowest among the similarly intelligent models it compared.
Vals AI found Ember finishes its tests in about 15 minutes, the fastest of the top 12 models it ranks.
Box, an early tester, used Ember to score contracts against a negotiation playbook and screen candidates against a detailed brief. Yashodha Bhavnani, Box’s VP of AI Products, said the model delivered “impressive accuracy, speed and token efficiency” on the company’s complex work evaluation.
Safety Checks and What’s Next
Before release, Meta evaluated Ember under its Superintelligence Scaling Framework, covering chemical and biological, cybersecurity, and loss-of-control risks, and found it stayed below the framework’s high-risk thresholds. Meta says the small model received the same core safety training and testing as its larger models, including checks on how it behaves when acting on a user’s behalf, such as resisting instructions hidden in web pages or tool outputs and asking for human input before high-impact actions.
Meta says it has more efficient models in training. Ember is the first of the specialized models Meta said it would build with Wonder Full when it switched Muse and Meta AI to Wonder-trained models last week.
