A Guth Labs publication ↗
Back to News

Meta targets cheaper AI agent work with Muse Ember

AI-generated conceptual editorial illustration of small glowing computation core routing documents and tools; branded guthlabs.ai.
AI-generated conceptual editorial illustration | guthlabs.ai

The new model matches Claude Haiku 5.5 on one independent index at a lower measured cost per task. Long prompts and different deployment platforms need a closer look.

Meta released Muse Ember on November 9, adding a lower-cost model for the repetitive work inside AI agents. It is available through the Meta Model API and Muse Code, priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens. Meta announcement

The company positions Ember as a worker alongside larger models: one system can plan a job while Ember processes documents, calls tools or handles coding steps. Meta advertises text, image, audio and video understanding, a one-million-token context window. Meta announcement

Independent testing gives the efficiency pitch some support. Artificial Analysis reports an Intelligence Index score of 36 at maximum reasoning effort, matching Claude Haiku 5.5 at that setting. The measured cost was about $0.14 per index task for Ember, compared with $0.41 for Haiku. Artificial Analysis

The difference is more subtle than a cheaper headline token rate. Artificial Analysis says the models have the same standard input and output prices, but Ember benefits from cheaper cached input and uses roughly one-third as many output tokens in its evaluations. Those workload-specific savings should not be treated as a promised discount on every application. Artificial Analysis

The results also reveal trade-offs. On FrontierCode 2.0, Ember scored 27.2%, behind Haiku’s 34.9%. On the knowledge-focused AA-Omniscience test, Ember declined to answer 64% of questions. That helped limit hallucinations, but also left its answer accuracy below similarly capable models in the comparison. A model that frequently abstains may need a stronger backup. Artificial Analysis

Long documents change the price calculation. Meta’s published rates rise to $0.30 per million input tokens and $1.50 per million output tokens when the prompt exceeds 200,000 input tokens. A large context window therefore does not mean the lowest advertised price applies across its full capacity. Meta developer guide

Deployment details differ too. Google Cloud lists Muse Ember 1.0 as a preview service, with text and image inputs and text output. Its documentation supports function calling and structured output, but does not list batch predictions or provisioned throughput as available. Teams evaluating the model through Google should use those documented limits rather than assume every launch capability is present. Google Cloud documentation

OpenRouter’s listing, by contrast, describes audio, video and PDF inputs alongside text and images. It also says the model is not currently available through its EU in-region routing. These differences make the chosen endpoint part of the purchasing and testing decision. OpenRouter

For developers, the useful test is a representative batch of actual work: measure successful completion, retry costs, human review and escalation to a larger model. Ember’s launch offers a promising way to reduce the expense of repeated agent steps. Whether that improves the overall workflow will depend on the steps it can complete reliably.

Sources and citations

Revision history

Published .