Meta's Muse Ember: the cheapest small model, built for agents that work all day
Frontier models have stopped competing on who is smarter and started competing on who is cheaper, because the real bill is not chat. It is agents calling tools, over and over.
Three key facts
Meta released Muse Ember on November 9, describing it as its smallest and most efficient model. The positioning is narrow on purpose: the execution layer for complex reasoning, handling tool calls, coding subtasks, routing, extraction and long document parsing. It costs $0.10 per million input tokens and $0.50 per million output tokens, available immediately on the Meta Model API.
Meta's comparison: at equivalent reasoning settings, Muse Ember matches Claude Haiku 5.5's best score at roughly a third of the cost, and beats every GPT-6 Luna setting. On Artificial Analysis data dated November 9, it scores 37.8 on end-to-end business workflows against 13.9 for Haiku 5.5 and 20.2 for Luna, and 40.9 on Terminal-Bench 4.0 against 39.2 and 16.4.
The same table shows where it loses. On FrontierCode 2.0, which judges real engineering tasks on mergeability, it scores 27.2, below Haiku 5.5's 34.9. The context window is 1 million tokens with up to 512K output tokens, enough to take a full codebase or document set in one call. Meta also says this is the first time one of its models sits on the efficiency frontier of the Artificial Analysis leaderboard: no model scores higher for less per task.
Wang Dou's Take
Read Meta's wording carefully. It says more intelligence per dollar, not more intelligence. In 2024 those were the same thing. In 2026 they are not.
The telling rows are the two agent-specific ones: OSWorld 2.1 for desktop use and AutomationBench for end-to-end business workflows. That is where the money actually burns. A customer service agent makes tens of thousands of tool calls a day at a few hundred tokens each. Shave a dime off the model and the bill halves. Meta cutting input to 10 cents per million tokens is really cutting the electricity bill of the agent economy.
Meta did not hide the coding gap. That is smart: it is not fighting for the job of writing hard code. It is taking the high-volume, low-complexity work where 80 percent is good enough. And about 90 percent of enterprise agent tasks are exactly that.
Buyers should change the question. Stop asking which model is smartest. Ask how many millions of tokens you burn every day, and how many of those cents were wasted.
Source: Meta AI Research Blog
