Meta Superintelligence Labs released Muse6.7 out of 10 Ember on Monday, November 9, which TechCrunch describes as the lab’s lightest and most cost-efficient model yet. Meta says Ember is “designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.”
Positioning and pricing
Ember is the fourth model under the Muse brand, following Muse Spark, its open-source counterpart Glimmer, and the Wonder frontier model. Meta pitches it as a partner to a larger model rather than a replacement: “Muse Wonder plans, Muse Ember handles the volume.” TechCrunch gives the example of a customer-support agent in which a heavier model sets the rules for an issue and a lighter model applies them across tens of thousands of cases.
Artificial Analysis lists pricing at $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. It notes those input and output rates match Claude Haiku 5.5 and GPT-6 Luna, while Ember’s cached input is 60% cheaper, which matters because cached input makes up most of the cost in agentic evaluations. Availability is through Meta’s first-party API and Muse Code. Meta says Ember takes text, images and audio, so agents can work from screenshots, scanned documents, voice memos and video without a separate model.
What the independent numbers show
Artificial Analysis reports that Ember at max effort scores 36 on its Intelligence Index, level with Claude Haiku 5.5, at $0.14 per task versus $0.41 for Haiku. It scores 8 points below Muse Wonder (44) at roughly one-fiftieth of the cost per task. On AutomationBench-AA it scores 68.1%, level with Muse Wonder.
Results are mixed elsewhere. Ember beats Haiku 5.5 on Terminal-Bench 4.0 (40.9% vs 32.8%) but trails it on FrontierCode 2.0 (27.2% vs 34.9%). Inside Muse Code it scores 46 on Artificial Analysis’s Coding Agent Index, ahead of Haiku 5.5 in Claude Code (37) and GPT-6 Luna in Codex (31).
Knowledge is the clear weakness. Ember’s AA-Omniscience accuracy is 21.1%, the lowest among comparable models Artificial Analysis tested, largely because it declines to answer 64% of questions. That restraint keeps its hallucination rate at 17.4%, about half of Haiku 5.5’s.
Why it matters
Agent workloads burn tokens on repetitive steps, so the cost of the cheap tier often decides what is affordable to run. Ember puts Meta in direct competition with Anthropic’s and OpenAI’s small models and with low-cost open-weight rivals such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Teams should note that the cost advantage was measured by Artificial Analysis on its own tasks, and Meta’s own comparison charts come from the same benchmark, so workload-specific testing is advisable before switching.