Meta launches Muse Ember for cheaper agent tasks
Meta launched Muse Ember on November 9 as its smallest model for repeated agent work, including tool calls, coding subtasks and document processing. It is available in Muse Code and through the Meta Model API as muse-ember-1.0. The starting price is $0.10 per million input tokens and $0.50 per million output tokens, but developers using long conversations need to watch a separate pricing threshold.
The useful question is where Ember can replace an expensive call without creating more retries or review work. Its strongest case is a narrowly defined job repeated many times. A large context window and an attractive benchmark score do not settle that decision on their own.
What Muse Ember supports
Meta’s model documentation lists text, image, video, audio and PDF inputs, with text output. The context window is 1,048,576 tokens and the maximum output is 524,288 tokens. Ember supports Responses, Chat Completions and Messages endpoints, along with tool calling, computer use, structured output and search grounding.
Reasoning effort ranges from minimal to max, with max as the default. That default deserves an explicit check when comparing speed or consumption against another model. Muse Code offers Ember through usage based billing and every subscription plan. On the API, it is available only in Standard mode, with no Contributor variant.
The price triples above 200K input tokens
For requests up to 200K input tokens, the official pricing page lists $0.10 input, $0.004 cached input and $0.50 output per million tokens. Above 200K, those rates triple to $0.30, $0.012 and $1.50.
Above 200K input tokens, the higher rates apply to the whole request, including output. Cached input and conversation history carried through a stored response count toward the threshold.
For example, 100K uncached input tokens plus 10K output tokens cost $0.015. At 300K input plus 10K output, the cost is $0.105. These are illustrative token costs. Search grounding adds $2.50 per 1,000 queries.
Standard API mode excludes submitted content from model training. Muse Code subscriptions have a separate Contributor setting that defaults to on, so check the product’s data controls before moving a sensitive workflow.
A practical budget should track cumulative input, cache use, output and retries. Pruning irrelevant history can matter as much as selecting the cheaper model, provided the removed context is not needed to finish the job accurately.
Good efficiency results, uneven coding strengths
Artificial Analysis reports an Intelligence Index score of 36 for Ember at max effort, with a cost of about $0.14 per task. Claude Haiku 5.5 at max also rounds to 36, at about $0.41 per task in that evaluation. These figures describe the evaluator’s workload and pricing assumptions, not a flat fee for any request.
The same analysis shows a mixed coding picture. Ember scores 40.9% on Terminal Bench 4.0 compared with Haiku’s 32.8% in those runs, but trails Haiku on FrontierCode 2.0, at 27.2% versus 34.9%. Artificial Analysis also reports limited factual knowledge: Ember answers many knowledge questions by declining rather than supplying an answer. That behavior can reduce incorrect answers while leaving more work for retrieval or a stronger model.
Vals AI’s separate evaluation places Ember at 60.48% on its index, ranking twelfth among 35 models at publication. It reports a $0.153 cost per test, the lowest on that index, while identifying weaker results in terminal work and tax tasks. Different test collections reveal different strengths. ByteForward has not independently reproduced these evaluations.
Read the benchmark setup before comparing scores
Meta’s evaluation report distinguishes internal comparisons from external results. Its internal tests generally use a common task framework, while outside coding figures preserve the reporting provider’s agent software. Meta’s Terminal Bench comparison uses Haiku’s self reported Claude Code result, which differs from Artificial Analysis’s comparison above.
Scoring rules matter too. Meta’s desktop evaluation includes partial credit, while a strict pass rate requires full completion. An average score should not be presented as the percentage of workflows that finished perfectly. A deployment test should specify the exact model, reasoning effort, tools, success criteria and token costs together.
Where Ember fits in an agent workflow
Meta’s model selection guide recommends Ember for classification, routing and bounded subtasks. It points developers toward Muse Spark or Muse Wonder for harder single request coding, knowledge heavy questions and visual coding. Its proposed split is straightforward: a larger model plans the work, while Ember carries out focused assignments.
For example, a team could test Ember on extracting a fixed set of fields from documents, then escalate incomplete or contradictory records. Measure the cost of accepted outputs after validation, including failed attempts and escalations. Keep approval steps around external actions. Lower inference prices do not make mistaken writes or messages cheaper to repair.
Our Muse Wonder access guide explains the larger model’s separate API and subscription options. For Ember, the sensible starting point is a representative batch of contained tasks with clear checks. Expand its role only when accuracy, latency and total cost hold up on that workload.
