Meta AI Research released Muse Ember on November 9, 2026 through the Meta Model API (as `muse-ember-1.0`) and Muse Code, priced at $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004 per million.

Meta describes it as the company's smallest model and its most efficient option for the repetitive layer of agent work: tool calls, coding subtasks, routing, extraction, and long-document parsing.

Ember offers a 1M-token context window and outputs up to 512K tokens, natively processes text, images, audio, and video, and is the first model trained using Muse Wonder Full.

The list prices match Claude Haiku 5.5 and GPT-6 Luna, so the claimed edge is in tokens burned. Per unite.ai, Meta's comparison with Muse Spark 1.3 shows Ember scoring 36 on the Artificial Analysis Intelligence Index versus 37 for Spark, at costs of $0.14 versus $2.39 per Artificial Analysis task.

Pasquale Pillitteri reports the fuller tier: Ember at 35.8 points and $0.136 per task, Haiku 5.5 at 36.2 and $0.411, and Luna at 27.8 and $0.152 — meaning Haiku spends roughly three times as much for a near-identical score.

Meta's intended division of labor is Wonder to plan, Spark for hard agentic workflows, and Ember to run repeatable versions at scale.

At maximum reasoning, Meta reports Ember beating both rivals on most agent evaluations: 37.8% on AutomationBench-AA business workflows (Haiku 13.9%, Luna 20.2%), 61.6% on the Linux subset of Agents' Last Exam (a Meta internal run), 40.9% on Terminal-Bench 4.0 (Haiku 39.2%, Luna 16.4%), and 46 on the Coding Agent Index-AA versus 37 and 31.

On OSWorld 2.1 partial credit it edges Haiku 69.5% to 68.9%.

The reversal is FrontierCode 2.0, which tests whether a repository maintainer would merge the work: Haiku wins at 34.9% against Ember's 27.2%, per Pasquale Pillitteri, who notes even open-weights DeepSeek scored 28.5% — a warning for anyone using Ember as a coding subagent.

Meta published preview testimonials from paying customers.

unite.ai reports Cloudflare CTO Dane Knecht said Ember ran up to 30% faster at the tail than the comparable model his company tested and handled multi-repository investigations well, while you.com inference lead Chak Pothina reported a 95th-percentile latency of 2.2 seconds, with Ember completing every coding quest at the lowest latency among models tested and producing 36–48% fewer output tokens.

Pasquale Pillitteri adds testimonials from Box, Cline, and Oracle, but cautions these are Meta-selected and Ember's per-task cost is a Meta estimate at its own list prices; Meta also does not state which countries can access the API.

Ember's launch lands the same day as a very different release: South Korea's ETRI announced TANGO 2, a no-code integrated development tool that automates the whole AI pipeline from training to deployment, with its core LLMOps solution published on GitHub as open source, per The Herald Business.

Where Meta keeps Ember's weights closed and sells by the token — only Muse Glimmer among the family is open-weights, as Pasquale Pillitteri notes — ETRI is handing developers the tooling itself, with pilots in shipbuilding, agriculture, and radiology at Seoul National University Hospital.

For working programmers, the practical takeaway is to benchmark Ember against their own workload, especially where mergeability, not cost, decides success.