Muse Ember, the lightweight model Meta Superintelligence Lab released on November 9, 2026, is now also available on Google Cloud and Oracle Cloud, extending its reach beyond Muse Code and the Meta Model API, where it launched under the model ID muse-ember-1.03,8. The model is priced at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. It is the first model Meta says it trained using its flagship Muse Wonder model, though the company did not specify which part of the workflow it automated with Muse Wonder4.
Meta said Ember delivers more intelligence per dollar than both Claude Haiku 5.5 and GPT-6 Luna, citing Artificial Analysis's composite benchmark score to support the claim1,2. At max reasoning, Meta puts Ember at about $0.14 per task on the Artificial Analysis Intelligence Index, compared with about $0.41 for Haiku and about $0.15 for Luna. Ember scores 36 on that index at max reasoning, matching Haiku's score of 36. Meta also claims Ember sits on the efficiency frontier of the full Artificial Analysis leaderboard, stating that no model scores higher for less per task7. At 10 cents per million input tokens, Ember is significantly cheaper than Haiku and Luna and cost-competitive with open-source alternatives DeepSeek-V4.1-Flash and GLM-5.3-Flash.
Built for repetitive agent work
Meta said Ember was "designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing". The intended deployment pattern pairs Ember with a heavier model such as Muse Wonder inside an agentic harness. In a customer-support scenario, for example, a larger model would establish rules for handling a particular issue, while Ember applies those rules across tens of thousands of cases. Meta described the division as: "Muse Wonder plans, Muse Ember handles the volume".
Ember supports a 1M-token context window and outputs of up to 512K tokens, allowing a full codebase or document set in a single call. The model accepts text, images, audio, and video as inputs. Meta also says Ember can operate desktop applications from screenshots.
Benchmark picture beyond cost
On OSWorld 2.1, Ember finished 33.8 percent of tasks outright, compared with 27.8 percent for Claude Haiku 5.5 and 17 percent for GPT-6 Luna. Ember scored 69.5 percent partial credit on the same benchmark, versus 68.9 percent for Haiku. On FrontierCode 2.0, Ember scored 27.2 percent, trailing Haiku's 34.9 percent but ahead of Muse Spark 1.3's 23.1 percent.
Meta's own comparison with Muse Spark 1.3 frames the trade-off within the Muse family: Spark 1.3 scores 37 on the Artificial Analysis Intelligence Index versus Ember's 36, but Spark's listed cost per task is $2.39 compared with Ember's $0.14. Meta also offers a Spark Contributor tier that matches Ember's uncached input price and charges $0.20 per million output tokens, though that tier uses requests to improve Meta's products.
Meta is keeping Ember's model weights private, consistent with its approach for previous Muse model releases. The model supports the Responses API.
ANALYSIS The pricing and positioning place Ember squarely in the lightweight-execution tier that Anthropic and OpenAI also serve with Haiku and Luna, respectively, but Meta's 10-cent-per-million-input-token price point forces a direct cost comparison that the company is clearly inviting. By designing Ember as a complement to Muse Wonder rather than a standalone frontier model, Meta is optimizing for the multi-model orchestration pattern in which a planner delegates high-volume subtasks to a cheaper executor.