Meta's Muse Ember targets cheap agent tool calls, tops Vals Index on cost

Muse Ember is Meta's entry in the small-model tier that now does most of the work inside agent systems. Meta positions it for tool calls, coding subtasks, request routing, structured extraction and long-document parsing, the high-frequency steps where cost per call dominates total spend.

The headline number comes from Vals AI, which Meta amplified: Ember is the cheapest model on the Vals Index at $0.15 per test while outperforming Anthropic's Haiku 5.5 and OpenAI's GPT-6 Luna. Dev.to commentary summarized it as Haiku-level intelligence at roughly one third of the cost. On HN, the launch drew 258 points, mostly from agent builders interested in cheap tool calling.

Small models are the battleground of the agent cost crisis. NVIDIA said this week that token usage is doubling monthly and pitched Switchyard routing to cut serving costs by about two-thirds; AWS's Swami Sivasubramanian highlighted Strands Decider matching an LLM at 98% less cost on a ticketing workload. Ember slots into that pattern: route cheap steps to a cheap model and reserve frontier models like Muse Wonder or Fable 5.2 for hard reasoning.

Caveats: a single index lead does not guarantee reliability on long tool-call chains, and Meta has not said whether Ember weights are open, which matters to developers who watched Llama's share of enterprise open-model spend fall to 7%. Watch for latency figures and availability on cloud platforms.