On the same day Anthropic made its frontier tier cheaper, Meta Superintelligence Labs went after the other end of the agent stack. Muse Ember, released November 9, 2026, is the fourth Muse model, after Spark (April), the open-weight Glimmer (August) and the frontier Wonder (October). Meta says it is the first model trained using Muse Wonder Full.
The pitch is narrow on purpose. Ember is built for “the work agents repeat all day”: tool calls, coding subtasks, routing, extraction and long document parsing. Its job is to be the cheap subagent in someone else’s harness.
Specs and pricing
| Muse Ember | |
|---|---|
| API model ID | muse-ember-1.0 |
| Input / cached / output (≤200K prompt) | $0.10 / $0.004 / $0.50 per MTok |
| Input / cached / output (>200K prompt) | $0.30 / $0.012 / $1.50 per MTok |
| Context / max output | 1M / 512K tokens |
| Modalities | Text, images, audio, video in |
| APIs | Meta Model API (Responses API supported), Muse Code, Google Cloud, Oracle Cloud |
The list price matches Claude Haiku 5.5’s entry rate. The cached-input price, $0.004 per million, is very low, which matters in agent loops that re-read the same prefix every turn.
Benchmarks: strong on agents, weaker on mergeable code
Meta’s table compares Ember at max reasoning effort with Claude Haiku 5.5 and GPT-6 Luna:
- Artificial Analysis Intelligence Index: 36, tied with Haiku 5.5 (Luna: 28), at $0.14 per task versus about $0.41 for Haiku at max.
- AutomationBench-AA (end-to-end business workflows): 37.8 vs 13.9 (Haiku) and 20.2 (Luna).
- Terminal-Bench 4.0: 40.9 vs 39.2, each model in its best published harness (mini-swe-agent for Ember).
- OSWorld 2.1 (partial credit): 69.5 vs 68.9.
- FrontierCode 2.0-AA: 27.2 vs 34.9 for Haiku. When the question is whether maintainers would merge the change, Haiku still wins.
Against Meta’s own mid-tier Muse Spark 1.3, Ember scores 36 to Spark’s 37 on the index, at $0.14 versus $2.39 per index task.
Partner quotes repeat the speed story. Cloudflare reports up to 30% faster tail latency than the comparable model it tested. You.com measured a p95 of 2.2 seconds on its coding quests, with 36–48% fewer output tokens than leading open-weight models.
Where it fits
Meta’s own framing is a two-tier harness: “Muse Wonder plans, Muse Ember handles the volume.” In practice that means:
- Explorer / reader subagents that grep, summarize files and parse logs, then return short notes to a more expensive planner.
- Extraction over long documents. The 1M context plus native vision means scanned PDFs, charts and screenshots don’t need a separate OCR model.
- High-volume classification and routing, where outputs can be checked cheaply.
It is a weaker fit for the final code edit on a hard change. The FrontierCode gap suggests keeping that step on a stronger model.
Caveats
- Most results come from Meta’s own evaluations or from Artificial Analysis runs at max reasoning effort. Test Ember at the effort setting you will actually run.
- Pricing triples above 200K prompt tokens, so a subagent that piles up context quietly gets 3x more expensive. Compact or reset before you cross that line.
- Ember is API-only. Unlike Glimmer, there are no open weights.
Bottom line
If your agent bill is mostly subagents reading files, Ember is worth a side-by-side test against Haiku 5.5 on your own traffic. Compare cost per completed task, not price per token.