Meta Debuts Muse Wonder AI Models as Investors Weigh Heavy Capex
Meta introduced Muse Wonder and Muse Wonder Full, calling them its most capable models, with API pricing and a focus on coding, agent tasks, and memory. The launch follows massive AI capital spending, including about $89 billion over the last four quarters. Analysts see limited upside and are waiting for usage or subscription figures to judge whether the models can offset those costs.
Meta rolled out two new models on Oct. 12: Muse Wonder, accessible via its API and Muse Code, and Muse Wonder Full, limited to Muse Code subscribers. Wonder costs $4 per million input tokens and $20 per million output tokens. Both target coding, agent work, and persistent memory.
The models were trained at Meta's Prometheus facility in Ohio. Meta's quarterly capex reached $30.12 billion in Q2 2026, up 82% year over year, with roughly $89 billion spent across the prior four quarters. Artificial Analysis rated Wonder Full 56, versus 48 for Muse Spark.
The launch could affect developers, businesses, and consumers as coding and agent tools become more capable and memory-aware. If such models reduce software costs, smaller firms may benefit, while some workers could face shifting job tasks. Subscription and API pricing may determine who can access advanced features, potentially widening gaps between large and small users. Meta's heavy spending also raises questions about whether ad-funded platforms or paying customers ultimately bear the cost, shaping how widely these tools spread.