AI
Meta rolls out Muse Ember, compact AI model to compete in efficiency-focused market
Meta's latest small language model targets speed and cost-consciousness, matching rivals Claude Haiku and GPT-6 Luna on performance while undercutting prices.
1/4
The New Stack
Meta introduced Muse Ember this week—a compact, multimodal model designed to handle routine AI tasks at minimal cost. The release reflects a broader industry strategy where tech companies maintain tiered model portfolios: nimble, inexpensive versions for focused work paired with heavier-duty models for complex reasoning.
Positioning against established rivals#
Muse Ember enters a crowded field occupied by OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5, both pitched for high-volume, straightforward jobs like reading documents, executing tests, and making targeted code changes. Meta positions Ember similarly, working alongside its flagship Muse Wonder Full for planning and orchestration—the model handles agentic subtasks while the larger variant manages higher-level strategy.
According to Meta's benchmarks on the Artificial Analysis Intelligence Index, Ember achieves around $0.14 per task at maximum reasoning settings, undercutting Haiku's $0.41 and staying competitive with Luna's $0.15. It matches Haiku's performance score at roughly one-third the cost, while slightly outpacing Luna for about a penny less per task.
Pricing and access#
Muse Ember supports a 1-million-token context window with up to 512K output tokens. Its pricing mirrors Luna and Haiku at standard short-context rates: $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. The model is live now through Meta's Code interface and the Meta Model API under the identifier muse-ember-1.0, with Responses API support included.
Developers can expect broader deployment soon across platforms like OpenRouter and other cloud services, following Meta's historical pattern with prior Muse releases. Model weights remain private.
Performance across benchmarks#
On Meta's Personal Superintelligence Index—which measures personal and knowledge-work capabilities—Ember leads both Haiku and Luna. However, it shows softer results on FrontierCode 2.0, a metric testing whether repository maintainers would accept an agent's code changes: Ember scores 27.2 percent versus 34.9 percent for Haiku and 28.5 percent for DeepSeek.
For desktop automation tasks measured by OSWorld 2.1, Ember achieved 69.5 percent partial credit (nearly identical to Haiku's 68.9 percent) but completed a higher percentage of tasks fully: 33.8 percent versus 27.8 percent for Haiku and 17 percent for Luna.

