Meta launches Muse Ember, a low-cost AI model for agent work
Meta's Muse Ember targets repetitive agent tasks with low token pricing, a million-token context window and benchmark claims that remain company-reported.
Meta has launched Muse Ember, a smaller AI model designed for the repetitive work that sits inside software agents: tool calls, coding subtasks, routing, extraction and long-document parsing. The model is available immediately through the Meta Model API as muse-ember-1.0 and through Muse Code. 1
Meta is positioning Ember around cost and throughput rather than frontier-scale general intelligence. The company lists pricing of $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. It also gives the model a 1-million-token context window and support for up to 512,000 output tokens.
The model is intended for workflows that break a large task into many smaller steps. Meta says Ember can pull records, call tools in sequence and update systems across apps, while its context window is large enough to take in a full codebase or document set in one call. The company describes the model as multimodal across text, image, audio and video.
What Meta claims about performance
Meta's announcement compares Muse Ember with Claude Haiku 5.5 and GPT-6 Luna across a series of evaluations. It reports a score of 37.8 on AutomationBench-AA at maximum reasoning, 40.9 on Terminal-Bench 4.0 and 61.6 on Agents' Last Exam. Meta also lists 80.6 on PerceptionBench with tools and 46.8 on its Personal Superintelligence Index. 1
Those figures are not independent tests. Meta says some results come from Artificial Analysis, while others are internal evaluations or reproduced from the company's own charts. The comparison table also shows Ember behind Claude Haiku 5.5 on FrontierCode 2.0-AA, so the launch's performance story is not a clean win across every workload.
Muse Ember is available through the Meta Model API and Muse Code, with support for the Responses API. Meta says it has also made the model available through Google Cloud and Oracle Cloud. The company describes it as the first model trained using Muse Wonder Full, and places it below Muse Wonder and alongside Muse Spark in its model portfolio. 1
Our opinion
Muse Ember makes more sense as infrastructure than as another model for people to chat with. Agent systems spend much of their time on small, repeated actions, so a cheap model with a large context window could be more useful than a headline-grabbing general-purpose model if it is reliable enough.
The important qualification is that Meta is asking readers to accept a lot of its own evaluation framing. The pricing is concrete, but the quality claims depend on selected benchmarks, internal tests and comparisons that may not reflect every real-world agent harness. Ember is worth watching for routine coding and document work, but its value will be decided by error rates, latency and how often developers need to hand a task back to a larger model.