Meta's smallest model is priced for the work agents repeat all day

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 9 November Meta Superintelligence Labs released Muse Ember, its smallest and most efficient model, built for "the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing". It is available on the Meta Model API and Muse Code as muse-ember-1.0 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens, with a one-million-token context window and up to 512,000 output tokens. It accepts text, images, audio and video. Meta says Muse Ember, Muse Spark and Muse Wonder are also available on Google Cloud and Oracle Cloud.

Meta's headline claim is cost per task. On the Artificial Analysis Intelligence Index, it says Ember matches Claude Haiku 5.5's best score at about a third of the cost and beats every setting of OpenAI's GPT-6 Luna. Several of its other comparisons are Meta's own evaluations, including a "Personal Superintelligence Index", and Haiku still leads on one coding test Meta reports. TechCrunch notes the price is competitive even with open-source rivals and describes the intended pattern, in Meta's words: "Muse Wonder plans, Muse Ember handles the volume." Meta says Ember is the first model trained using Muse Wonder Full, received the same core safety training, stayed below the high-risk thresholds of its scaling framework, and was evaluated on whether it "seeks human input before high-impact actions".

Why it matters

Agent systems are settling into tiers. An expensive model makes the plan; a cheap one carries it out thousands of times. The economics are compelling, and they move risk in a direction many firms have not noticed: the model that touches the most records, calls the most tools and runs with the least supervision is the cheapest one.

In a tiered agent, the cheap model does most of the touching. It deserves most of the testing.

For a bank or insurer, that means validation effort should follow volume, not prestige. If a small model extracts fields from 50,000 claim forms a day, its error rate on those forms matters more than the big model's benchmark score. Test the executor on your own documents, measure what happens when the planner gives it an ambiguous instruction, and make sure it stops and asks before high-impact actions, as Meta says it was evaluated to do, rather than guessing.

What I'd watch

• Independent results for Muse Ember on Artificial Analysis and on extraction tasks with real business documents.
• Whether Meta publishes Ember's behaviour on its 'seek human input' evaluation in detail.
• How quickly rival labs cut prices for their smallest models in response.