Meta Releases Its Smallest Model Muse Ember, Targeting Low-Cost Agent and Coding Tasks

Meta AI Research introduced Muse Ember on November 9, 2026, describing it as the company’s smallest model and its most efficient option for repeated agent and coding work. It became available the same day through Muse Code and the Meta Model API, priced at $0.10 per million input tokens and $0.50 per million output tokens.

The target workload is the repetitive layer of agent activity: tool calls, coding subtasks, routing, extraction, and long-document parsing. Meta says agents spend most of their tokens on these small, structured steps. Ember can retrieve records, call tools in sequence, and update systems across applications, allowing it to execute a business workflow from start to finish.

Pricing and Muse Family Roles

Cached input costs $0.004 per million tokens. The model is exposed as muse-ember-1.0 and supports the Responses API. Meta says its 1M-token context window can accommodate a full codebase or document set in one call, while the model supports outputs of up to 512K tokens.

Meta’s comparison with Muse Spark 1.3 frames the trade-off: Ember scores 36 on the Artificial Analysis Intelligence Index, versus 37 for Spark 1.3, while the listed costs per Artificial Analysis task are $0.14 and $2.39, respectively. Ember is also the first model trained using Muse Wonder Full. Meta’s intended division is for Wonder to plan the hardest problems, Spark to handle coding and agentic workflows, and Ember to perform repeatable versions of those tasks at scale.

Agent and Coding Evaluation Results

Meta says Ember matches Claude Haiku 5.5’s best Intelligence Index result at roughly one-third of its cost per task and beats GPT-6 Luna at every reasoning setting, with no model on the full Artificial Analysis leaderboard scoring higher for less per task. Ember’s costs in that comparison are estimates at Meta’s list pricing.

At maximum reasoning, Meta reports the following evaluation results:

Evaluation Muse Ember Claude Haiku 5.5 GPT-6 Luna

AutomationBench-AA, business workflows 37.8% 13.9% 20.2%

Agents’ Last Exam, Linux subset 61.6% 57.0% 57.1%

Terminal-Bench 4.0 40.9% 39.2% 16.4%

Coding Agent Index-AA 46 37 31

FrontierCode 2.0-AA 27.2% 34.9% 25.2%

OSWorld 2.1, partial credit 69.5% 68.9% 50.5%

AutomationBench-AA used strict pass@1 scoring, and Terminal-Bench placed each model in its best published harness. Agents’ Last Exam and OSWorld were Meta internal evaluations. Haiku’s FrontierCode result qualifies the broader coding pitch: that benchmark evaluates real engineering tasks on mergeability, and Haiku leads it.

Ember also natively processes text, images, audio, and video. An agent can therefore work from screenshots, charts, scanned documents, voice memos, meeting recordings, or video without calling a separate model. These inputs extend the execution model to desktop computer use and visually detailed documents alongside text-based coding.

Safety and Early Partner Testing

Meta evaluated Ember under its Superintelligence Scaling Framework for chemical and biological, cybersecurity, and loss-of-control risks. The company reports that the model stayed below the framework’s high-risk thresholds across the categories assessed and received the same core safety training and testing. Agent-behavior evaluations covered respect for user intent, sensitive-information handling, resistance to instructions hidden in web pages, documents or tool outputs, and seeking human input before high-impact actions.

Meta also published early-testing accounts from Cloudflare and You.com. Cloudflare CTO Dane Knecht said Ember ran up to 30% faster at the tail than the comparable model his company tested and handled investigations spanning multiple repositories well. You.com inference lead Chak Pothina reported 95th-percentile latency of 2.2 seconds, with Ember completing every coding quest at the lowest latency among the models tested. Pothina also said Ember produced 36–48% fewer output tokens during those tests.

For developers, Ember creates a lower-cost execution layer that can sit beneath Muse Wonder’s planning and reduce the number of repetitive calls sent to a more expensive model. Meta’s numbers support that division of labor, while the FrontierCode result marks its limit: efficiency and strong agent scores do not guarantee superiority on every engineering task.