Meta Muse Ember: 1M Context, 512K Output, a Third of Haiku's Cost

Meta Superintelligence Labs released Muse Ember on November 9, 2026, as the small-model execution tier for repetitive agent work: tool calls, routing, extraction, coding subtasks, and long-document parsing. Meta says Ember is the first model trained using Muse Wonder Full, its largest model, and slots it below Muse Wonder for planning. The intended deployment is a two-tier stack—"Muse Wonder plans, Muse Ember handles the volume"—with Ember carrying high-volume subtasks.

Meta reports a 1M-token context window and up to 512K output tokens. MarkTechPost's launch comparison puts the rival output caps at 128K each for Claude Haiku 5.5 and GPT-6 Luna, which makes Ember's output window four times larger. A single call can load a full codebase or document set instead of chunking across requests, which is the tradeoff discussed in Long Context or Retrieval? Compare Cost, Quality and Update Work. Ember natively accepts text, images, audio, and video; MarkTechPost notes neither Haiku 5.5 nor Luna handles audio or video natively. Weights are closed. Access is limited to the Meta Model API and Muse Code as muse-ember-1.0, with no self-hosted, quantized, or on-device option.

All benchmark figures are as published by Meta at max reasoning effort; some runs come from Artificial Analysis, others from Meta's internal suite. Ember scores 37.8% on AutomationBench-AA, 657 Zapier-based business workflows, against 20.2% for GPT-6 Luna and 13.9% for Claude Haiku 5.5. On Agents' Last Exam, Meta's Linux-subset internal evaluation, Ember scores 61.6% versus Luna's 57.1% and Haiku's 57.0%. The AA Coding Agent Index gives Ember 46 to Haiku's 37 and Luna's 31. FrontierCode 2.0-AA, judged on mergeability, is the clear weak spot: Haiku 5.5 scores 34.9% to Ember's 27.2%, with Luna at 25.2%.

On the AA Intelligence Index, Meta reports Ember scores 36 at max, tied with Haiku 5.5 and ahead of Luna at 28. The cost gap is the more consequential result: Meta reports Ember costs $0.136 per index task at max versus Haiku 5.5's $0.411 and Luna's $0.152—roughly a third of Haiku's cost per task.

Ember lists at $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004 per million tokens. MarkTechPost reports Haiku 5.5 and GPT-6 Luna carry the same base $0.10/$0.50 rates up to their long-context thresholds, then Haiku applies surcharges above 100K prompt tokens and Luna doubles input rates above 272K input tokens. Meta lists no long-context tier for Ember. Since base token prices match, the cost-per-task gap has to come from Ember consuming fewer tokens per completed task, as MarkTechPost's comparison points out.

Against AI Mastery's pricing data, Ember's $0.10/$0.50 undercuts Mistral Small 4 ($0.15/$0.60) and is far below Claude Sonnet 5 ($2/$10) and Gemini 3.8 Flash ($0.75/$3.75). Those comparisons matter mainly for teams deciding whether to route high-volume agent work to a small model rather than a larger general-purpose one.

AI Mastery analysis

The cost-per-task framing is the useful signal, but engineers should pressure-test it on their own task distribution. The AA Intelligence Index is a composite, and the advantage can widen or narrow by subtask mix. FrontierCode 2.0-AA is the main caveat: Haiku 5.5's 34.9% against Ember's 27.2% shows mergeable code quality and token efficiency are not the same objective. Teams using a small model to close pull requests rather than pass isolated tests should weight that result heavily.

Closed weights are the other architectural limit. Teams with data-boundary or on-prem requirements cannot use Ember at all, regardless of benchmark profile, because every call crosses Meta's hosted inference. That shared surface—not the model itself—is where agent security failures tend to occur.

Meta's two-tier framing is coherent for high-volume document and agent pipelines where an orchestrator already exists. The 512K output ceiling and flat long-context pricing strengthen that fit. Whether the benchmark lead survives production depends on how closely AutomationBench-AA's 657 Zapier workflows match the actual tasks in a given pipeline, a caveat that applies to any evaluation-driven model choice.

Sources

• Primary: Introducing Our Smallest, Most Efficient Model: Muse Ember
• Reported by: MarkTechPost: Meta Releases Muse Ember, a 1M-Context Model for High-Volume Agent Work
• Competitor prices: AI Mastery token calculator, figures as of 2026-10-22