AI and security briefsModels

Is Meta's Muse Ember worth testing for routing, extraction and tool calls?

By Brandon Da Costa, founder of Digiton Dynamics

Published 9 November 2026Updated 9 November 2026Editorial policy

Test it on your cheapest, most repeated steps. Meta released Muse Ember on 9 November 2026 for tool calls, routing, extraction and coding subtasks, at $0.10 and $0.50 per million tokens. Meta does not state its size or licence, and its own table shows it behind Claude Haiku 5.5 on one coding test.

Who should act: Teams that run agents with many small repeated steps, and pay a large model to route, extract or format tool calls that a cheaper model might handle.

What changed

Meta released Muse Ember on 9 November 2026 and calls it its smallest and most efficient model, built for the work agents repeat all day (Meta AI blog, 9 November). It is on the Meta Model API and in Muse Code as muse-ember-1.0 and supports the Responses API. Meta lists $0.10 per million input tokens, $0.004 for cached input and $0.50 for output, with a 1M token context window and up to 512K output tokens. The post gives no parameter count, no licence and no word on open weights.

Meta says it is natively multimodal across text, images, audio and video and is the first model trained using Muse Wonder Full. Its advice for hard work is to pair it with Muse Wonder, so Wonder plans and Ember handles the volume (Meta AI blog, 9 November).

These are Meta's figures at max reasoning, compared with Claude Haiku 5.5 and GPT-6 Luna (Meta AI blog, 9 November):

• Artificial Analysis Intelligence Index: Ember 36, Haiku 36. Meta says Ember reaches that at about a third of Haiku's cost per task.
• FrontierCode 2.0-AA: Ember 27.2%, Haiku 34.9%, Luna 25.2%.
• Terminal-Bench 4.0: Ember 40.9%, Haiku 39.2%, Luna 16.4%, each in its own best harness.
• AutomationBench-AA: Ember 37.8%, Luna 20.2%, Haiku 13.9%.

Sources disagree in places. Artificial Analysis's own article gives Ember 68.1% on AutomationBench-AA and Haiku 32.8% on Terminal-Bench 4.0. It also says Ember's knowledge is its main weakness, with 21.1% accuracy on AA-Omniscience against 34.7% for Haiku, because it declines 64% of questions (Artificial Analysis, 9 November).

Who it affects

• Teams whose largest model handles routing, extraction or tool-call formatting. Those steps run most often, so they carry most of the bill.
• Haiku 5.5 and GPT-6 Luna users. The listed token prices match, but Meta's cached input price is lower and Artificial Analysis says Ember used about a third as many output tokens as Haiku.
• Anyone who needs open weights or a clear licence. The post states neither, so do not plan self-hosting on it.
• Anyone who needs recall of facts. Artificial Analysis reports a weak result, so keep fact lookup on a larger model or a retrieval step.

What to do now

1. List the steps in your agent and mark the ones that are short, structured and repeated: classify the request, pull fields into a schema, fill tool arguments. Leave planning and open-ended writing on the large model.
2. Pull 100 real inputs per step from your logs and write the correct output for each. Run the large model and Ember over the same set three times.
3. Score what a pipeline needs: schema-valid rate, field-level exact match, correct tool and arguments, and how often it declines. Add up cost per accepted result, not per call.

   jq -s 'group_by(.model)|map({model:.[0].model,n:length,valid:(map(select(.schema_ok))|length),usd:(map(.usd)|add)})' results.jsonl

4. Check that the large model and Ember were run with the same prompt, effort and output limit. Meta compared models at max reasoning, and at lower effort the gap may differ.
5. Route in two stages. Send a step to Ember, validate the output, and fall back to the large model when validation fails. Raise the share of traffic only while the fallback rate stays low.

What to watch

• A licence and weights statement. Meta has not given one.
• Independent runs on your task type. Meta ran most of this table itself.
• Decline behaviour on extraction. A refusal to guess is useful on a schema field and costly when the answer was in the document.
• Meta's remark that it has even more efficient models in training.

Questions people ask

Digiton's production review checks which model each agent step calls and what it costs per accepted result, and the checks are on the vibe coded app audit page.

How this brief was made: researched from the sources below, drafted with AI assistance and checked against each source on the dates shown. See the editorial policy.

Sources

1. Meta AI blog post introducing Muse Ember, 9 November 2026, read 9 November 2026. https://research.meta.ai/blog/muse-ember
2. Artificial Analysis article on Muse Ember and the Intelligence Index, 9 November 2026, read 9 November 2026. https://artificialanalysis.ai/articles/meta-s-muse-ember-is-the-cheapest-model-scoring-36-on-the-intelligence-index
3. TechCrunch report on Muse Ember, 9 November 2026, read 9 November 2026. https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/

More briefs

How these briefs are sourced and corrected. All AI and security briefs