Meta releases "Muse Ember," its smallest and most efficient AI model, at $0.10 per million input tokens, says it beats Haiku 5.5 on performance per dollar; also offered on Google Cloud

The smallest model, for agents' repetitive work

Meta released the AI model "Muse Ember" in an official blog post dated Nov. 9. The company described it as its "smallest, most efficient model" and said it is aimed at carrying out tasks involving complex reasoning and at work that AI agents repeat all day. It gave examples such as tool calls, fine-grained coding tasks, task routing, information extraction and reading long documents.

Meta said Muse Ember completes its model lineup, dividing roles with "Muse Wonder" for the hardest problems and "Muse Spark" for coding and agentic work. It described Muse Ember as a model that handles large volumes of recurring coding and agentic work, running at about $0.14 per task in Artificial Analysis' evaluation. It is the first model trained using "Muse Wonder Full," the company said. Meta also described a combination for hard problems in which Muse Wonder makes the plan and Muse Ember handles the high-volume work.

The model natively understands text, images, audio and video, and can handle screenshots, charts, scanned documents, recordings and videos without a separate model, the company said. Before release, Meta evaluated chemical and biological, cybersecurity and loss-of-control risks under its safety framework, the "Meta Superintelligence Scaling Framework," and said the model fell below the high-risk thresholds in every area evaluated.

API pricing of $0.10 input and $0.50 output

Muse Ember became available the same day under the name muse-ember-1.0 in the coding tool "Muse Code" and the "Meta Model API." Pricing per million tokens is $0.10 for input, $0.004 for cached input and $0.50 for output. These prices apply to prompts up to 200,000 tokens; for longer prompts, they rise to $0.30 for input, $0.012 for cached input and $1.50 for output.

The context window is 1 million tokens, and a single output can be up to 512,000 tokens (512K). Meta said an entire codebase or a batch of documents can be loaded in a single call. It also supports the "Responses API."

In Meta's comparison with the higher-end Muse Spark 1.3, Muse Ember is priced at $0.10 input and $0.50 output against Spark 1.3's $1.25 input and $4.25 output. On Artificial Analysis' overall "Intelligence Index," Muse Ember scored 36 against Spark 1.3's 37, and its cost per task on the index's evaluations was $0.14 against Spark 1.3's $2.39, the company said. All figures are at the deepest reasoning setting (max).

Meta says performance per dollar beats Haiku 5.5 and Luna

Meta said Muse Ember has more intelligence per dollar than Anthropic's "Claude Haiku 5.5" and OpenAI's "GPT-6 Luna" at every reasoning setting. It said Muse Ember matches Haiku's best score at about one-third of the cost and beats Luna at every setting. It also sits on the efficiency frontier across all models listed by Artificial Analysis, with no model having a lower cost per task and a higher score, the company said.

The chart it offered as evidence plots Artificial Analysis' overall index against the cost per task on the index's evaluations, for each reasoning setting. In the chart's description, Meta called Muse Ember's cost an estimate at proposed pricing, while a note below the chart described it as cost at list price. A footnote said cost per task was calculated using each model's list price.

Haiku 5.5 scores higher on some development evaluations

In the Artificial Analysis evaluations Meta published (all with reasoning at max, as of Nov. 9), Muse Ember scored 37.8% on AutomationBench-AA (the share strictly correct in one attempt), which tests completing business workflows automatically from start to finish, above Haiku 5.5's 13.9% and GPT-6 Luna's 20.2%. However, on FrontierCode 2.0-AA, which grades whether real development work can be merged, Muse Ember scored 27.2% against Haiku 5.5's 34.9%. On GDP.pdf-AA, which covers professional work using long documents, Muse Ember and Luna tied at 22.8%.

In Meta's internal evaluations, Muse Ember scored 69.5% on OSWorld 2.1 (scored with partial credit), which tests operating computer applications, against 68.9% for Haiku 5.5 and 50.5% for Luna. On Agents' Last Exam (a subset of Linux problems), which tests completing agentic tasks, it scored 61.6%, above Haiku 5.5's 57.0% and Luna's 57.1%. On Terminal-Bench 4.0, which measures coding in the terminal and was compared using the best agent harness each model has published, Muse Ember scored 40.9%, Haiku 5.5 39.2% and Luna 16.4%. The harnesses were mini-swe-agent for Muse Ember, Claude Code for Haiku 5.5 and Codex for Luna.

Also offered on Google Cloud, at the same prices as Meta's API

Meta Chief Executive Officer Mark Zuckerberg posted on X on Nov. 9 (6:49 a.m. Nov. 10 Japan time) that Muse Ember, Muse Spark 1.3 and Muse Wonder are now available on Google Cloud. Sundar Pichai, CEO of Alphabet and Google, replied about 50 minutes later, "welcome to the model garden."

Google Cloud's product page says the three models can be deployed from the "Google Cloud model garden" and used as serverless APIs. For Meta's models, customers can choose the pricing method, such as dedicated endpoints or pay-as-you-go, it says.

In Google Cloud's partner model price list, Muse Ember 1.0 costs $0.10 per million input tokens, $0.50 per million output tokens and $0.004 for cache-hit input when input is 200,000 tokens or less. Above 200,000 tokens, it costs $0.30 for input, $1.50 for output and $0.012 for cache-hit input. The figures are the same as Meta's API pricing. In the same list, Muse Spark 1.3 costs $1.25 input and $4.25 output, and Muse Wonder 1.0 costs $4.00 input and $20.00 output, with these two priced the same regardless of length.

Meta's official blog says the three models are also available on Oracle Cloud. The official pages checked do not state the regions where they are available on Google Cloud.