Meta releases Muse Ember, claiming it beats Haiku 5.5 and GPT-6 Luna on cost per task

Meta launched Muse Ember, the smallest and cheapest model in its commercial Muse lineup, alongside Muse Spark 1.3. Ember has a 1-million-token context window and targets high-throughput autonomous agent tasks such as business agents and deep data retrieval, where inference cost matters more than peak capability.

The headline claim comes from Vals AI, which Meta's AI account 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. Meta frames it as winning on intelligence per dollar, the metric that matters for agents making thousands of calls.

The small-model tier is crowded. Gemini 4 Carbon's aggressive pricing, Strands Decider's 98% savings claim at Amazon, NVIDIA's Switchyard router and cheap Chinese open-weight models all compete for the same cost-sensitive workloads. Ember gives Meta a credible commercial entry after it shifted from open Llama releases to closed Muse models.

Context matters: Ember shipped the same week Meta cut about 8,000 jobs while budgeting $135–145B in AI capex, so investors want proof that the spending pays off. The Vals result is a single third-party index. Watch whether developers actually adopt Ember, given Meta's smaller API footprint compared with OpenAI, Anthropic and Google.