Meta Superintelligence Lab releases Muse Ember, a low-cost lightweight model priced at $0.10 per million input tokens

Meta has released Muse Ember, the fourth and lightest model in its Muse family, from its Meta Superintelligence Lab. It costs $0.10 per million input tokens. Meta says it handles text, images and audio and is aimed at high-volume agent work such as tool calls, coding subtasks, routing, extraction and long-document parsing. It is meant to pair with the larger Muse Wonder model, which plans while Ember does the volume work. The output token price and full specifications were not in the coverage we saw.

Developers and enterprises running agent systems that make many small model calls now have another low-cost option to test, which could cut bills if quality holds. The pricing puts pressure on small models from Anthropic and OpenAI and on cheap open-source models such as DeepSeek V4.1 Flash and GLM 5.3 Flash. The real saving cannot be judged yet: output tokens are often the bigger cost in agent work, and that price is missing. Teams should treat this as a candidate for a trial, not a reason to switch.

The claim that Ember is "cheaper and smarter per task" than Haiku and Luna comes from Meta's own use of an Artificial Analysis comparison, so it has not been independently checked. The coverage is secondary: TechCrunch quotes Meta's announcement in detail but does not link to it. "Per task" is also a vendor-chosen measure, and the real cost of an agent job depends on output length, retries and failure rates, none of which have been shown.

1. 01Meta's published output token price and full spec sheet (context window, audio and image limits, rate limits), which will show the true cost per task compared with Haiku, Luna, DeepSeek V4.1 Flash and GLM 5.3 Flash.
2. 02Independent benchmark results and the Artificial Analysis listing for Ember over the next two to four weeks, especially on tool-calling accuracy and long-document parsing rather than Meta's chosen comparisons.
3. 03Price responses or new small-model releases from Anthropic, OpenAI, DeepSeek or Zhipu, and any developer reports on whether Ember paired with Muse Wonder cuts real agent costs in production.