Meta releases Muse Ember for high-volume agent work
Meta released Muse Ember on November 9 for repeated agent work, including tool calls, coding subtasks, routing and document processing. The model gives developers another execution option beneath larger planning models, with independent evaluation results and multiple hosting routes already available.
Spec sheet
- Model
- Muse Ember 1.0
- Independent index
- 36 at max reasoning, Artificial Analysis
- Index task cost
- $0.14 at max reasoning; benchmark-specific
- Meta context window
- 1 million tokens
- Google Cloud status
- Preview, global endpoint
A smaller executor beneath the planner
Meta positions Ember as the model for work an agent performs repeatedly, rather than its hardest planning decisions. Its launch recommends pairing Ember with Muse Wonder, assigning planning to the larger model and volume work to the smaller one. Meta says Ember is the first model trained using Muse Wonder Full. The company lists availability through its own Model API and Muse Code, with text, image, audio and video understanding.
Independent results separate task cost from token prices
Artificial Analysis reports an Intelligence Index score of 36 for Ember at maximum reasoning effort, matching Claude Haiku 5.5 at its maximum setting. Its measured cost per index task is $0.14 for Ember and $0.41 for Haiku. Those are task costs from that evaluation, rather than prices for a fixed quantity of tokens. Coding results also vary by test. Ember scores 27.2% on FrontierCode 2.0, below Haiku's 34.9%, so the tied aggregate score does not mean identical strengths across engineering work.
Gateway access exposes the surrounding application controls
Vercel added Ember to AI Gateway on November 9 under the identifier meta/muse-ember. Its announcement lists text, image and video inputs, a one-million-token context window and reasoning settings from minimal through max. Developers can reach the model through the AI SDK or supported Chat Completions, Responses and Anthropic Messages interfaces. The integration lets an application feed tool results into later steps, keeping the model inside a continuing workflow. Gateway also supplies usage and cost tracking, routing and configurable retry or failover behavior. These are service features around the model: choosing Ember does not itself specify how an application will handle a failed request. Vercel says it applies provider pricing without an inference markup or a separate platform fee. This separates the gateway charge from the model's token rates. Existing applications can keep a supported interface while changing their selected model.
The Google-hosted version has a narrower input contract
Google Cloud's November 9 model page lists Muse Ember 1.0 as a preview service on a global endpoint. It accepts text and images and returns text, with maximum input of 1,048,576 tokens and maximum output of 524,288 tokens. The page lists standard pay-as-you-go and fixed quota support, while excluding batch predictions and Provisioned Throughput. An audio workflow needs a different supported endpoint.
Claim check
- Confirmed Artificial Analysis reports Ember at 36 and $0.14 per index task at max reasoning, compared with Haiku 5.5 at 36 and $0.41. [2]
- Company claim Meta describes Ember as its smallest and most efficient model and says it was trained using Muse Wonder Full. [1]
- Confirmed Google Cloud lists Ember as preview, with text and image input; batch predictions and Provisioned Throughput are unsupported. [4]
- Unclear The published benchmark does not establish the cost, latency or completion rate of a particular production workflow.
Sources
- Introducing Our Smallest, Most Efficient Model: Muse Ember · Meta · 9 Nov
- Meta's newest model Muse Ember scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 (max) at a third of the Cost per Task · Artificial Analysis · 9 Nov
- Muse Ember now available on AI Gateway · Vercel · 9 Nov
- Muse Ember 1.0 · Google Cloud · 9 Nov