Key takeaways

  • Meta released Muse Ember on November 9 on the Meta Model API and Muse Code as muse-ember-1.0.
  • It costs $0.10 per million input tokens and $0.50 per million output tokens, with a 1M-token context window.
  • Meta says it beats Haiku 5.5 and GPT-6 Luna on intelligence per dollar, but Haiku scores higher on FrontierCode 2.0-AA.

Meta released Muse Ember on November 9, a small model built for agent work, and priced it at $0.10 per million input tokens and $0.50 per million output tokens. Meta says Muse Ember gives more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, according to Meta's announcement. It is available now on the Meta Model API and in Muse Code as muse-ember-1.0.

Ember is the fourth model in the Muse family, after Spark, the open-source Glimmer and the frontier model Wonder. It matters because the cheap, fast tier is where agents spend most of their tokens, and Meta is now selling into it at a price below Haiku's headline numbers.

What Muse Ember is for

Meta describes Ember as its smallest and most cost-efficient model, aimed at the repeatable work agents do all day: tool calls, coding subtasks, routing, extraction and long document parsing. TechCrunch reports that it is trained on text, images and audio, so an agent can read screenshots, charts and scanned documents.

Meta's framing is that Wonder plans and Ember handles the volume. In its example, Wonder sets the rules for customer-support cases and Ember applies them across many tickets.

Price and limits

Muse Ember
Input $0.10 per million tokens
Cached input $0.004 per million tokens
Output $0.50 per million tokens
Context window 1 million tokens
Maximum output 512,000 tokens

It supports the Responses API. OpenRouter lists the same $0.10 and $0.50 prices for Meta as the provider, and showed the model as newly listed on November 9.

What the benchmarks show

Most of Meta's numbers come from Artificial Analysis, an independent benchmarking firm, but Meta chose which results to publish. On the Artificial Analysis Intelligence Index, Meta reports Ember at 36 at maximum reasoning, level with Haiku 5.5 at 36 and ahead of GPT-6 Luna at 28. The cost per task is the stronger claim: $0.14 for Ember, against $2.39 for Muse Spark 1.3, which scores 37.

Benchmark (maximum reasoning) Ember Haiku 5.5 GPT-6 Luna
AutomationBench-AA 37.8% 13.9% 20.2%
Terminal-Bench 4.0 40.9% 39.2% 16.4%
OSWorld 2.1 (binary) 33.8% 27.8% 17.0%
Coding Agent Index-AA 46 37 31
FrontierCode 2.0-AA 27.2% 34.9% 25.2%

Ember does not win everywhere. Haiku 5.5 scores higher on FrontierCode 2.0-AA, and Ember only ties Luna on GDP.pdf-AA. Meta's lead on Terminal-Bench is also not like for like, because each model ran in its own best harness: mini-swe-agent for Ember, Claude Code for Haiku and Codex for Luna. Agents' Last Exam, the Personal Superintelligence Index and part of OSWorld are Meta's own evaluations.

What we don't know yet

The partner quotes are early and unverified. Cloudflare says Ember is up to 30% faster at the tail than a comparable model, and You.com reports a p95 latency of 2.2 seconds. Meta has not said whether Ember's weights will be open, which was the case for Glimmer, Spark's counterpart. Meta says Ember stayed below its high-risk thresholds for chemical and biological, cybersecurity and loss-of-control risks.

For anyone choosing a model for high-volume agent work, the next step is to run Ember on your own tasks. Vendor charts rarely match a real workload. For more on Meta's other Muse products, see our comparison of Meta Muse and ChatGPT.

Sources

  1. Our smallest, most efficient model: Muse EmberMeta
  2. Meta releases Muse Ember, a new lightweight model for agent executionTechCrunch
  3. Meta: Muse EmberOpenRouter

Add ClaudeGPT Si as a preferred source on Google