A small, efficient model core.
By Mara Quinn | November 10, 2026
Meta Launches Muse Ember to Cut the Cost of Agentic Work
The expensive part of running an AI agent often is not the reasoning. It is the scaffolding - the constant stream of tool calls, routing choices, coding subtasks, data extractions, and document parses that hold the pipeline together. Meta is betting that those tasks do not need a large model, and on November 9 it released Muse Ember to prove it.
Ember is the smallest and most efficient model Meta has built to date, according to the company. It is designed specifically for the repetitive work that agents do routinely, and Meta is pricing it accordingly. At $0.10 per million input tokens and $0.50 per million output tokens on the Meta Model API, it undercuts the cost of comparable small models by a margin that the company says third-party benchmark data supports.
A Smaller Model With a Large Context Window
Despite its positioning as a lightweight model, Ember carries a 1 million token context window. That capacity matters for the long-document parsing use case Meta lists among Ember's target tasks. Processing a lengthy contract, a research corpus, or a multi-file codebase requires holding a lot of text in memory at once, and Ember's window is large enough to handle most practical document workloads without chunking.
Ember is also multimodal. It accepts text, images, audio, and video as inputs, putting it beyond the text-only tier that most small models occupy. That range gives developers more flexibility in what they pipe through it.
Access is API-only. Meta is not offering Ember as a downloadable model, which means teams that need on-premise deployment or the ability to fine-tune model weights will need to look elsewhere.
A caching feature reduces costs further for inputs that repeat across requests. When caching is active, cached input tokens drop to $0.0004 per million. For agent loops that repeatedly pass the same system prompt or reference document, that rate makes the economics more favorable still.
Benchmark Claims: Tied on Quality, Cheaper Per Task
Meta's core performance claim rests on the Artificial Analysis Intelligence Index, which aggregates nine benchmarks into a single composite score. Ember and Anthropic's Claude Haiku 5.5 both scored 36 on that index. The difference, according to Meta's cited data from Artificial Analysis, shows up in cost: completing a comparable task with Ember runs roughly $0.14, versus roughly $0.41 with Haiku at maximum reasoning mode. That is about one-third the cost per task.
Meta also claims that Ember outperformed Haiku and OpenAI's GPT-6 Luna across 7 of the 9 constituent benchmarks. The largest recorded lead appeared in business task automation, measured under a benchmark Meta identifies as AutomationBench-AA, where Ember posted a 17 percent advantage over those competitors.
One important caveat: all benchmark numbers cited here come from Meta or from Artificial Analysis as cited by Meta. No independent, hands-on verification of Ember's performance had been published at the time of the announcement. These figures should be treated as vendor claims until third-party testing appears.
Built on Meta's Own Training Infrastructure
Ember is the first model Meta trained using Muse Wonder, its flagship system. The move indicates that Meta is running its model development pipeline through its most capable model to produce smaller, specialized outputs - a pattern that other labs have used to distill high-quality capability into compact form. Meta has not specified which part of the training workflow Wonder handled, so the degree of automation involved is not clear from the announcement.
Early Partners
Meta named five integration partners ready to work with Ember at launch: Box, Cline, Oracle, Cloudflare, and You.com. The company did not detail the specific use cases for each partner in the initial announcement.
What Remains Open
Ember's positioning puts it in a narrow slot: capable enough to handle multimodal, million-token workloads, cheap enough to run at high volume in agent pipelines. The parity score with Haiku on the composite index, combined with a cost-per-task roughly one-third lower, is the most striking data point in the release. Whether independent testing confirms that profile is the question that will determine how widely developers adopt it.
The API-only constraint limits its reach. Teams that need portability or model control will not find it here. But for developers building on Meta's infrastructure who want to reduce the marginal cost of agentic operations, Ember represents a specific answer to a specific problem.
