← All articles
Meta launches Muse Ember, its smallest model for repetitive agent work
AI News

Meta launches Muse Ember, its smallest model for repetitive agent work

Muse Ember is Meta's smallest model, built for tool calls, coding subtasks and document parsing at $0.10 per million input tokens.

Meta Superintelligence Lab launched Muse Ember on November 9, 2026, the smallest and most cost-efficient model in its Muse family. It is built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing. The model is live today on the Meta Model API as muse-ember-1.0 and in every Muse Code plan.

Official Muse Ember launch banner from Meta
Official Muse Ember launch banner. Image: Meta.

1. What Muse Ember is

Muse Ember is an execution model, not a planner. Meta's own framing: "Muse Wonder plans, Muse Ember handles the volume." In a customer support agent, a larger model like Muse Wonder sets the rules for handling an issue, then Ember applies them across tens of thousands of cases.

This is the fourth model in the Muse family. Muse Spark shipped in April 2026, Muse Glimmer in August, Muse Wonder in October, and Ember in November. Meta says Ember is the first model trained using Muse Wonder Full.

2. Pricing and size

Pricing is $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K tokens, according to Meta's developer blog. For longer prompts, the price rises to $0.30 input and $1.50 output. The context window is 1 million tokens, with up to 512K output tokens.

Meta claims Ember matches Claude Haiku 5.5's best score at about a third of the cost and beats every GPT-6 Luna setting, based on Artificial Analysis data from November 9, 2026. Its cost sits at about $0.14 per Artificial Analysis task. TechCrunch separately reported that Ember is cost-competitive with open source models like DeepSeek V4.1 Flash and GLM 5.3 Flash. These figures are Meta's claims and media coverage, not independently verified benchmarks.

3. Multimodal without a second model

Ember natively understands text, images, audio and video. Agents can read screenshots, charts and scanned documents, or work from voice memos and meeting recordings, without a separate model. That matters for mixed-format agent work, for example: read an invoice image and return the invoice number, issue date and total.

4. Partner test results

Meta's early partners shared their numbers. You.com said Ember finished every coding task in their testing with the lowest latency of any model they tested, a p95 of 2.2 seconds, roughly 4x to 13x faster than leading open-weight models, using 36 to 48 percent fewer output tokens. Cline called it purpose-built for subagent work like document analysis.

In Meta's own example, one workflow captioned 10,000 photos, renamed them and sorted them into albums for $1.59 in total, of which only $0.50 was the Muse Ember cost itself.

5. Safety evaluated before release

Meta says Ember was evaluated under the Meta Superintelligence Scaling Framework before launch, covering chemical, biological, cybersecurity and loss-of-control risks. All assessed categories stayed below the framework's high-risk thresholds. Meta also tested how the model behaves when acting on a user's behalf, including resistance to hidden instructions in web pages and tool outputs.

What it means for you

Small-model pricing has now fallen to the 10-cents-per-million-token mark, the same tier as Haiku 5.5 and GPT-6 Luna. The differentiator is no longer list price but real cost per task.

  1. Build two-tier agents. Use a large model to plan and a cheap model like Ember to execute at volume. That is the official pattern Meta demonstrated.
  2. Check cost per task, not cost per token. The $0.14 per task figure from Artificial Analysis matters more to an agent budget than a price list.
  3. Match work to modality. If your agent processes invoices, screenshots or voice memos, one cheap multimodal model can replace a chain of two.
  4. Benchmark for yourself. The "one third of the cost" claim comes from Meta. Test on your own workload before migrating.
  5. Track your AI stack. For tool comparisons, see AI Tools.

Sources