Meta's Muse Ember isn't cheaper per token. It's cheaper per task

• Muse Ember: $0.10 per million input tokens, $0.004 cached input, $0.50 output (Meta).
• Claude Haiku 5.5: $0.10 input, $0.01 cache hits, $0.50 output for prompts up to 100,000 tokens (Anthropic).
• GPT-6 Luna: $0.10 input, $0.01 cached input, $0.50 output at standard short-context rates (OpenAI).

The difference is in how the bill adds up. Artificial Analysis, the independent benchmark firm Meta cites, says Ember's cached input, "most of the cost in our agentic evaluations," is 60% cheaper, and that Ember uses roughly a third as many output tokens as Haiku 5.5.

Cost per task. On the Artificial Analysis Intelligence Index, Ember (max reasoning) scores 36, the same rounded score as Claude Haiku 5.5 (max), at about $0.14 per task versus $0.41 for Haiku. GPT-6 Luna's best setting scores 28. Meta's own Muse Spark 1.3 scores 37 but costs $2.39 per task.

Where it falls short.

• Mergeable code. On FrontierCode 2.0, which judges whether a maintainer would merge an agent's change, Ember scores 27.2% against 34.9% for Haiku 5.5.
• Knowledge. Artificial Analysis says knowledge is Ember's main weakness: 21.1% accuracy on its AA-Omniscience test versus 34.7% for Haiku 5.5. Ember declines to answer 64% of those questions, which keeps its hallucination rate at 17.4%, about half of Haiku's.
• Benchmarks differ by harness. Meta reports Terminal-Bench 4.0 at 40.9% for Ember versus 39.2% for Haiku, using each model's best published harness. In Artificial Analysis's own run, Haiku scored 32.8%. Several other numbers on Meta's chart (Agents' Last Exam, OSWorld 2.1, PerceptionBench, and Meta's Personal Superintelligence Index) are Meta's internal evaluations.
• Closed weights. Meta offers Ember through its Model API and Muse Code (OpenRouter and Vercel's AI Gateway also listed it on launch day). Per The New Stack, Meta is keeping the weights private.

Our take. Ember looks built for one job: a subagent that runs many small, repeated steps against the same cached context. That is where the cheaper cache and shorter answers pay off. It is not the model to trust with final code review or with facts it has to know. If your agent bill is mostly cached input, run Ember on your own workload and compare cost per finished task, not price per token.

Disclosure: Not tested by MODEL. Not sponsored. No affiliate links.

MODEL Editors