Meta shipped its cheapest model on November 9. Muse Ember costs $0.10 per million input tokens, carries a 1M-token context window, and is purpose-built for the work that blows agent budgets — tool calls, extraction, routing, document parsing, and repetitive coding subtasks. At roughly one-third the cost of Claude Haiku 5.5 for the same intelligence level on agentic tasks, it enters one of the most competitive model slots in the market right now.
Ember’s Job in the Muse Stack #
Meta’s model lineup has a clear division of labor. Muse Wonder handles the hard stuff — multi-hour planning runs, complex orchestration, the main reasoning loop. Muse Spark owns coding-heavy workflows. Muse Ember runs the volume: the thousands of small, repeated subtasks that make up most of what a production agent actually does.
Meta’s own documentation makes the mapping explicit. For teams already using an OpenAI-compatible harness, the recommendation is to alias your haiku or sub-agent model slot to muse-ember-1.0, Spark to muse-spark-1.3, and Wonder to muse-wonder-1.0. The API is OpenAI-compatible — switching an existing call is just a model field change.
response = client.chat.completions.create(
model="muse-ember-1.0",
messages=[{"role": "user", "content": "Extract all invoice fields from this document."}]
)
Ember is available today on the Meta Model API, Muse Code, and Cloudflare Workers AI.
Benchmarks: Strong on Agent Tasks, Weaker on Pure Coding #
The benchmark story is mixed in a useful way. On agentic and automation tasks, Ember beats both Claude Haiku 5.5 and GPT-6 Luna by a meaningful margin. On AutomationBench-AA, Ember scores 37.8% against Haiku’s 13.9% and Luna’s 20.2%. On Terminal-Bench 4.0, it posts 40.9% versus Haiku’s 39.2% and Luna’s 16.4%. The Coding Agent Index has Ember at 46, Haiku at 37, Luna at 31.
The flip side: for straight-line, single-shot coding questions, Haiku 5.5 still wins. On FrontierCode 2.0, Haiku scores 34.9% versus Ember’s 27.2%. Meta is upfront about this — if the task is a hard coding problem, a knowledge-intensive question, or visual coding work, the docs say to reach for Muse Spark or Muse Wonder instead.
The honest summary: Ember is better than Haiku 5.5 at agent execution and costs a third of the price. It is worse at the kind of coding that requires deep reasoning in one pass.
The 200K Billing Cliff: Understand This Before You Ship #
The headline price is $0.10/M input, $0.50/M output. That rate holds up to 200K input tokens per request. Past that threshold, the entire request reprices to $0.30/M input and $1.50/M output — not just the excess tokens, the whole thing. One token over the line and that call costs three times as much.
This matters more than it looks. Conversation history counts toward the input token total. Teams building chat-style agents or passing full repository context in every call will hit the cliff without noticing. The practical fix is to keep request context under 200K — chunk documents, use caching, or summarize history before it compounds. The official pricing page has the full breakdown.
The 1M-token context window is real and genuinely useful for batch processing large files. Just do the math before treating it as a free upgrade.
The Bottom Line #
Muse Ember is a credible option for the sub-agent and background task layer of an AI system. The cost advantage over Haiku is substantial, the agentic benchmarks are genuinely better, and the OpenAI-compatible API makes it low-friction to evaluate. The limits are clear: it is not the right model for hard reasoning in a single call, and the 200K billing cliff will surprise teams that do not plan for it.
If you are already running a tiered model strategy, Ember is worth evaluating for the execution layer. If you are running a single-model setup and your tasks are reasoning-heavy, it is not a straight swap for Haiku.