On November 9, Meta introduced Muse Ember, the smallest and cheapest model in its Muse family. It is not meant as an all-rounder but for the routine work AI agents repeat thousands of times: tool calls, small coding tasks, routing, data extraction and reading through long documents. Pricing is $0.10 per million input tokens and $0.50 per million output tokens. According to Meta, Muse Ember matches Claude Haiku 5.5 at about a third of the cost per task – a claim based on Artificial Analysis measurements and the vendor’s own internal tests.
Muse Ember: what Meta announced
After Muse Wonder for the hardest problems and Muse Spark for coding and agentic workflows, Ember fills the gap at the bottom. In its announcement on research.meta.ai, Meta describes the model as an execution model: a larger model plans, Ember handles the volume. Meta says Ember is the first model trained using Muse Wonder Full – when it introduced Wonder, Meta had said it would use it to develop specialized models.
Key specs according to the Model API documentation:
- Model ID:
muse-ember-1.0(“Standard mode” only, no Contributor variant) - Inputs: text, image, video, audio and PDF; output: text
- Context window: 1,048,576 tokens, max output: 524,288 tokens
- Reasoning effort: from “minimal” to “max”, default “max”
- Endpoints: Responses, Chat Completions and Messages; according to Meta, OpenAI-compatible clients only need a different
base_url - Features: computer use, image understanding, tool calling, structured output and search grounding
Ember is available through the Meta Model API and in every plan of the Muse Code coding tool. Meta says Muse Ember, Muse Spark and Muse Wonder are also available on Google Cloud and Oracle Cloud. The announcement makes no mention of open weights – unlike the larger Muse Spark Open.
Pricing: cheap up to 200,000 tokens
Meta tiers its prices by prompt length:
- Up to 200,000 input tokens: $0.10 per million input tokens, $0.004 for cached input, $0.50 per million output tokens.
- Above 200,000 input tokens: $0.30, $0.012 and $1.50 respectively – three times as much in each case.
For comparison, Meta lists Muse Spark 1.3 at $1.25 (input) and $4.25 (output). Per task in the Artificial Analysis Intelligence Index, Ember costs around $0.14 according to Meta, Spark 1.3 around $2.39 – at an index score of 36 versus 37. In its developer blog, Meta gives “about $0.13” per task for Ember; both figures come from the same measurement and differ only in rounding or estimation basis.
Benchmarks: strong on agents, weaker on mergeable code
Meta compares Ember at maximum reasoning effort with Claude Haiku 5.5 and GPT-6 Luna. Some numbers come from Artificial Analysis (as of November 9), others from Meta’s own evaluation suite. All figures are vendor claims:
- Artificial Analysis Intelligence Index: Ember 36, Haiku 5.5 36, Luna 28.
- AutomationBench-AA (end-to-end business workflows): Ember 37.8%, Luna 20.2%, Haiku 13.9%.
- Coding Agent Index-AA: Ember 46, Haiku 37, Luna 31.
- Terminal-Bench 4.0: Ember 40.9%, Haiku 39.2%, Luna 16.4% – each in its best published harness, so not under identical conditions.
- FrontierCode 2.0-AA (mergeability on real engineering tasks): Haiku 34.9%, Ember 27.2%, Luna 25.2% – Ember clearly trails here.
- OSWorld 2.1 (internal, partial credit): Ember 69.5%, Haiku 68.9%, Luna 50.5%.
- Agents’ Last Exam (internal, Linux subset): 61.6% vs. 57.0% (Haiku) and 57.1% (Luna).
- GDP.pdf-AA (long documents): Ember and Luna 22.8% each, Haiku 20.8%.
The core value claim rests on cost per task: at max reasoning, Ember scores 35.8 index points for an estimated $0.136 per task, Haiku 5.5 scores 36.2 points for $0.411. That is where Meta’s “about a third of the cost” comes from. Note that Meta labels Ember’s costs as estimates based on list pricing, while the competitors’ values are measured by Artificial Analysis.
Partner quotes and a worked example
Meta quotes several test partners. Cloudflare reports up to 30 percent faster responses “at the tail” than a comparable model, You.com a p95 latency of 2.2 seconds in its own coding tests. In the developer blog, Meta shows an example: Muse Wonder plans porting 16 lexers of the Python library Pygments to Rust while 16 Ember agents work in parallel, each in its own Git worktree. The 16 ports reportedly finished within about six minutes of each other. This has not been independently verified; the example and its prompts are in the Meta Model Cookbook.
Safety and limits
Before release, Meta says it evaluated Ember under its “Meta Superintelligence Scaling Framework” for chemical and biological risks, cyber capabilities and loss of control. The model stayed below the high-risk thresholds in all categories assessed. Meta also tested whether Ember respects user intent, handles sensitive information appropriately, resists instructions hidden in web pages or tool outputs, and asks for human input before high-impact actions. The announcement does not include a detailed system card with individual results; Meta points to an evaluation report.
Practical limits follow from the positioning: for hard single-shot tasks, knowledge-heavy questions or visual coding, Meta explicitly recommends Muse Spark or Muse Wonder. And anyone using long contexts pays triple above 200,000 input tokens. In agent loops with a growing history, that threshold can be reached sooner than expected.
Our take
Muse Ember targets a market that is growing fast with agent workflows: small, quick models acting as subagents under the control of a large model. Anthropic follows the same division of labor – Haiku is a common low-cost subagent, and with Claude Fable 5.2 the company presented experiments in which one model coordinates many agents. Meta is mainly betting on price – and on a model family in which planners and workers share the same request structure, so switching often means changing only the model ID.
Whether the benchmarks hold up remains to be seen in independent tests. Notably, Ember trails Haiku clearly on FrontierCode, which is exactly where mergeable code for real engineering tasks matters. For routing, extraction and tool chains, on the other hand, the model could be a low-cost option. Teams already working with Muse Wonder or Spark can test Ember as a worker model with little effort.
