Tool Suggestions
Meta Muse Ember: The $0.10 Model Built for the Agent Work You Repeat All Day
Meta launched Muse Ember on November 9, its smallest and most efficient Muse model, built for the work AI agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing. It costs $0.10 per million input tokens and $0.50 per million output tokens, the same sticker price as Claude Haiku 5.5 and OpenAI's GPT-6 Luna, and it runs through the Meta Model API and Muse Code under the model ID muse-ember-1.0.
Small model, big context
The small label refers to operational footprint, not context capacity. Muse Ember has a 1,048,576-token context window and supports up to 524,288 output tokens, so it can hold long documents and long agent sessions in memory. Meta says it is the first model trained using Muse Wonder Full, the top tier of its Muse Wonder model, and it is the fourth model under the Muse brand, following Spark in April, Glimmer in August and Wonder in October.
The intended setup is two models, not one. Meta's documentation recommends pairing Ember with Muse Spark or Muse Wonder: the larger model handles planning and reasoning, while Ember takes on the repeated execution. Mark Zuckerberg described Ember as our smallest model that delivers higher intelligence per dollar than others in its class.
The price, and the 200,000-token tripwire
The headline price is $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. The detail that matters is the threshold: requests with more than 200,000 input tokens move to a higher tier, where the whole request bills at three times the base rate, $0.30 input, $0.012 cached and $1.50 output. A long-context request therefore costs triple, not just the part above the threshold, so checking request size before sending is part of using Ember well.
Ember is available on the Meta Model API in Standard mode only, with no Contributor-mode variant. Meta has kept the model's weights and parameter count private, so there is no open-weight version to run yourself.
Developers call it as muse-ember-1.0 through Meta's Responses, Chat Completions and Messages endpoints, and it is also inside Muse Code, Meta's coding tool, where it can take the small jobs inside a larger coding session. The launch material mentions no waitlist: the model is available immediately, in Standard mode only, which means anyone with a Meta Model API account can try it on a real workload today.
What Meta's benchmarks say, and what they cost
Every score in this section is Meta's own, and Meta notes that its cost chart figures are estimates at proposed pricing. At maximum reasoning, Meta puts Ember at 36 on the Artificial Analysis Intelligence Index, against 37 for Muse Spark 1.3, and at $0.14 per Artificial Analysis task against $2.39 for Spark 1.3 on the same comparison. On AutomationBench-AA at maximum reasoning, Ember scores 37.8%, against 13.9% for Claude Haiku 5.5 and 20.2% for GPT-6 Luna.
The about a third of the cost claim needs reading carefully. It refers to per-task cost on Artificial Analysis benchmarks, not to the sticker price of each token: Ember's sticker price is identical to Haiku 5.5 and GPT-6 Luna at $0.10 input and $0.50 output. What Meta is claiming is that Ember finishes the benchmark tasks for less money overall, because it needs fewer tokens or fewer retries. Nobody had independently reproduced those numbers at launch.
Where Ember fits, and where it does not
Meta's portfolio line is explicit: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable coding and agentic tasks at scale. Ember is the executor, not the planner, and it is priced for volume: the model only saves money if the work really is repetitive. For a one-off hard problem, the larger models are still the right call, and for anything above 200,000 input tokens the triple rate eats into the savings.
The honest test is a small one. Take a job your agent repeats, a routing decision, a file extraction, a batch of similar coding subtasks, and run it through Ember alongside the model you use now. Compare the bill and the failure rate, not just the sticker price. If Ember holds up, it is the cheapest way to run that job; if it does not, the savings were never real.
The comparison that matters most is with the two models at the same sticker price. Claude Haiku 5.5 and GPT-6 Luna both charge $0.10 input and $0.50 output, so Ember does not win on price per token. It wins, if Meta's numbers hold, on cost per finished task: fewer tokens, fewer retries, a lower bill for the same job. That is a claim about the model's efficiency, not its price list, and it is the number to test first before moving any real workload onto it.