Muse Ember Launches: 17x Cheaper, Commenters Zero In on $0.14 per Task

Alexandr Wang announced Muse Ember on X. He calls it the smallest model to date, offering intelligence close to Muse Spark 1.3 (max) at a 17x lower price, by AA's measure. It is also the first model trained with Muse Wonder Full. It is now available to try in the API and in muse code.

Price was the first point amplified in the replies. Viber said he read the $0.14-per-task figure several times; it basically means you can throw every little idea in and run it. Elliott Michael Smith's response was: once something is 17x cheaper, you no longer have to be so polite when asking for another revision. fw.estee said that being 17x cheaper suggests the old models were making a bit too much money before. The $0.14-per-task figure came from an online reply; the official tweet did not provide a full price list, but it shifted developers' psychological anchor. When trial-and-error costs are low, retries become more frequent.

The performance section was relayed by Justin. He calls Muse Ember Meta's smallest model to date, built for agent work, with a 1M context window. According to the Meta chart he posted, Muse Ember compared with Haiku 5.5 and GPT-6 Luna: AutomationBench 37.8 vs. 13.9 and 20.2; Coding Agent Index 46 vs. 37 and 31; Terminal-Bench 4.0 40.9 vs. 39.2 and 16.4. A 1M context window is more practical for agent tasks that run across codebases, read long documents, and call tools in sequence. Justin also posted a video of a one-shot website generated with the workofart skill; the front-end result is impressive for a small, fast, cheap model.

These numbers come from a vendor chart; independent replication has not yet appeared. What is clear is that the goal for small models has been spelled out: they can run tasks continuously and make developers willing to run them more often. Answering nicely in chat is no longer the only metric. The competitors on the benchmark table are Haiku 5.5 and GPT-6 Luna, and the battle has spread to the agent toolchain.

Another thread in the replies was open source. Ankit Kumar asked directly, 'open source?' Vladimir and am.will pressed for open weights. The official materials did not answer. Others asked whether Muse is already in use, when it will come to Europe and India, and whether there is a research paper. Prabhath is waiting for a comparison with Haiku and Luna. Arv misread the name as Musk Amber. Some posted memes, and someone said 'meta coming for all the labs.' The questions under the launch image were very specific.

The visible change is the pricing unit. In the past, model pricing was usually discussed per token. The Muse Ember replies pulled the conversation to $0.14 per task. Per-task pricing is closer to developers' real decisions: is this job worth having an agent run once, and is the cost of rerunning it after a failure acceptable? If small models can push the cost per task down to nearly free, many edge-case tasks that previously were not worth automating will start getting tried.

What to watch next is independent testing, open weights, and availability by region. The bold red on the launch image only shows the vendor's confidence; developers will vote with their own task lists.