Meta Prices Muse Ember at $0.10 per Million Tokens and Pitches It as the Subagent That Does an Agent's Busywork

Meta has released Muse Ember, a small, low-cost model aimed at the repetitive steps that make up most of an AI agent’s work: tool calls, coding subtasks, routing, extraction and long-document parsing. Meta Superintelligence Labs announced the model on November 9, 2026. It is available on the Meta Model API and in Muse Code as muse-ember-1.0, and Meta says Muse Ember, Muse Spark and Muse Wonder are now also offered on Google Cloud and Oracle Cloud.

Meta calls Ember its “smallest and most efficient model,” though it does not disclose a parameter count. It is the first model Meta says it trained using Muse Wonder Full, the larger system it introduced earlier as a base for building specialized models.

Priced for volume

Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens. Above that, the rates rise to $0.30, $0.012 and $1.50 respectively. The model has a 1-million-token context window, can return up to 512,000 output tokens, and supports the Responses API.

For comparison, Meta lists Muse Spark 1.3 at $1.25 per million input tokens and $4.25 per million output tokens. On the Artificial Analysis Intelligence Index, Meta reports Ember at 36 against 37 for Spark 1.3, with a cost per index task of about $0.14 for Ember versus $2.39 for Spark.

The subagent pitch

The release is framed around how agent systems actually spend tokens. “Agents spend most of their tokens on many small, structured steps, and Muse Ember is built for exactly that,” Meta writes. It describes Ember as the model for the coding subtasks “an agent sends off hundreds of times a session: reading files, running tests, making targeted fixes,” and suggests pairing it with the larger model: “Muse Wonder plans, Muse Ember handles the volume.”

Ember accepts text, images, audio and video natively, which Meta says lets an agent read screenshots, charts and scanned documents, or work from meeting recordings, without routing them to a separate model. It can also operate a computer from screenshots.

Partner statements in the launch post follow the same theme. Cline founder Saoud Rizwan said the model is “purpose-built for, and excels at, subagent work like coding and document analysis.” Cloudflare CTO Dane Knecht said Ember “ran up to 30% faster at the tail than the comparable model we tested.” You.com’s inference lead Chak Pothina reported a p95 latency of 2.2 seconds on its coding tasks and 36–48% fewer output tokens than leading open-weight models. Box and Oracle also contributed early-testing comments.

How Meta’s numbers compare

Meta compares Ember with Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna, all at maximum reasoning effort. Its table shows Ember at 37.8% on AutomationBench-AA, against 13.9% for Haiku and 20.2% for Luna; 40.9% on Terminal-Bench 4.0, narrowly ahead of Haiku’s 39.2%; and 46 on the Coding Agent Index-AA, against 37 and 31. Haiku 5.5 leads on FrontierCode 2.0-AA, a test of whether code changes would be merged, at 34.9% to Ember’s 27.2%. On the overall Artificial Analysis index, Ember and Haiku tie at 36.

Meta says Ember “matches Haiku’s best score at about a third of the cost.” Several figures come from different sources, and the post labels each one. Artificial Analysis results are dated November 9, 2026, with Ember’s cost per task calculated at its list pricing. Agents’ Last Exam (61.6% for Ember), OSWorld 2.1 and PerceptionBench are Meta’s internal evaluations, and the “Personal Superintelligence Index” is an internal composite of eight tests. Terminal-Bench 4.0 used a different harness for each model: mini-swe-agent for Ember, Claude Code for Haiku and Codex for Luna.

Below Meta’s risk thresholds, agent-behavior scores unpublished

Meta says Ember was evaluated under its Meta Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks and stayed below the framework’s high-risk thresholds in every category assessed. “Though it’s a small model, it received the same core safety training and testing,” the company writes. Meta also tested agent behavior, including whether Ember respects user intent, handles sensitive data appropriately, resists instructions hidden in web pages, documents or tool outputs, and asks for human input before high-impact actions. The launch post does not publish scores for those agent-behavior tests.

Ember fills out a three-tier lineup: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Ember for repeatable work at scale. Meta says it has “even more efficient models in training.” The launch post lists only hosted availability and does not mention downloadable weights.