Meta has released Muse Ember, a small reasoning model priced for the repetitive jobs that AI agents run thousands of times: calling tools, fixing code, sorting requests, pulling data out of documents. Meta AI Research announced it in a blog post on 9 November. The price is 10 cents for each million tokens read and 50 cents for each million written, and Meta says the model is live now.

Meta’s headline claim is about cost against quality. Using the Artificial Analysis leaderboard, the company says Ember at its top setting scores 35.8 on the Intelligence Index for $0.136 per task. Claude Haiku 5.5, Anthropic’s small model, reaches 36.2 at its top setting but costs $0.411 per task. That is where Meta’s “about a third of the cost” comes from. Against GPT-6 Luna, OpenAI’s small model, Meta says Ember wins at every setting; Luna’s best result in the chart is 27.8 at $0.152.

The agent benchmarks carry the sales pitch. On AutomationBench-AA, a test of end-to-end business workflows, Meta reports 37.8 for Ember, 20.2 for Luna, and 13.9 for Haiku. On Terminal-Bench 4.0, which tests coding in a command line, it reports 40.9 against Haiku’s 39.2 and Luna’s 16.4. Haiku still wins one coding test Meta lists: FrontierCode 2.0-AA, where it scores 34.9 to Ember’s 27.2. Several of the other results, including Agents’ Last Exam and OSWorld 2.1, are marked in the post as Meta’s internal evaluations, so outsiders cannot yet check them.

The model reads text, images, audio, and video, and Meta says it can operate a computer from screenshots. It takes a 1M-token context window, enough for a whole codebase or document set in one request, and can write up to 512K tokens back. Cached input is cheaper still, at $0.004 per million tokens.

Access is hosted only, according to Meta’s post. Ember runs on the Meta Model API under the name muse-ember-1.0, and in Muse Code. The post adds that the Muse family can also be rented from Google Cloud and Oracle Cloud. It mentions no downloadable weights and no open license, so developers should treat this as a paid API product, not an open model.

Box, the enterprise content company, is the one outside voice in the post. Yashodha Bhavnani, Box’s VP of AI Products, said that on Box’s own Complex Work Eval the model “delivers impressive accuracy, speed and token efficiency,” and described it scoring contracts against a negotiation playbook clause by clause. That is a vendor partner praising a vendor’s product in the vendor’s own announcement.

Meta also describes how the models fit together. Ember is the first model trained using Muse Wonder Full, the larger system Meta reserves for the hardest problems. Meta pitches Muse Wonder as the planner and Ember as the worker that handles volume. Muse Spark 1.3, Meta’s model for coding and agent work, scores 37 on the same index against Ember’s 36, but at $2.39 per task versus $0.14, with output priced at $4.25 per million tokens against Ember’s $0.50.

On safety, Meta says it tested Ember under its own Superintelligence Scaling Framework, which covers weapons-related science, hacking, and AI systems slipping out of human control. Meta reports the model stayed under the framework’s top risk tier. It also says it checked whether Ember can be hijacked by text planted in a web page or file, a real attack on any agent that reads untrusted content.

One footnote on the cost comparison: the chart labels Ember’s per-task costs as estimates at proposed pricing. The post gives no independent test of Ember beyond the leaderboard figures Meta chose to reproduce.

Anyone paying for a small model to run an agent’s bulk steps now has a $0.14-per-task alternative to price against Haiku 5.5 and Luna. The test that settles it is your own workflow, run on all three, before the next contract renewal.

Reported by Meta AI Research on 9 November 2026.