Meta Launches Muse Ember Lightweight Model with Input Price as Low as $0.10 per Million Tokens

Meta's Superintelligence Lab has released Muse Ember, its lightest and most cost-efficient model to date, designed for high-frequency agent tasks such as tool calling, coding subtasks, routing, information extraction, and long document parsing. Input pricing is as low as $0.10 per million tokens, with output at $0.50 per million tokens, and it supports a 1 million token context with native multimodal understanding. Meta claims it outperforms Anthropic Haiku and OpenAI Luna on cost and intelligence based on the Artificial Analysis composite benchmark, and it competes on price with open-source rivals like DeepSeek V4.1 Flash and GLM 5.3 Flash. Ember is positioned as the agent execution layer paired with the larger Muse Wonder model, where the heavy model plans and the light model executes high-frequency tasks.

Cross-validation from multiple sources shows Muse Ember scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 (max), but at a cost of only $0.14 per task, about one-third of the latter. On the Code Arena WebDev leaderboard, it ranks first among small models with a score of 1608, surpassing Claude Haiku 5.5 and GPT-6 Luna, with a median cost per task of just $0.03, placing it on the Pareto frontier. It scores 68.1% on AutomationBench-AA, on par with Meta's flagship Muse Wonder, though its overall intelligence lags the flagship by 8 points at about one-fiftieth the cost. However, on the real-world coding benchmark FrontierCode 2.0-AA, it trails Haiku with 27.2% versus 34.9%, indicating clear capability boundaries.

The release of Muse Ember marks a new phase in the efficiency race among small models, offering near-frontier intelligence at extremely low cost, which could significantly lower the barrier for deploying agent applications and accelerate automation of high-frequency tasks. Meanwhile, Meta's tiered inference architecture (Muse Wonder for planning, Ember for execution) demonstrates the potential of model composition, likely prompting the industry to reassess cost-effectiveness and intensifying price competition with open-source lightweight models.