Meta has launched Muse Spark, the first flagship model to emerge from the Superintelligence Labs division it built around chief AI officer Alexandr Wang, in the company's most concrete answer yet to a year of questions about whether its multibillion-dollar bet on artificial intelligence is producing results. The company said the model delivers competitive performance across multimodal perception, reasoning, health and agentic tasks while running at a fraction of the compute cost of the mid-size variant of its older Llama 4 family.
The cost claim is the heart of the pitch. The economics of frontier AI have come to hinge less on raw capability than on the price of delivering it at scale, and Meta is positioning Muse Spark as a model that narrows the gap with OpenAI and Google not by outspending them on inference but by undercutting them on it. If the efficiency figures hold up under independent testing, they would mark a meaningful shift for a company whose earlier models were praised for openness but rarely for frontier-level performance.
The launch is also a referendum on Wang himself. Meta restructured its sprawling AI efforts into Superintelligence Labs and handed him the top role after an expensive talent campaign that reshaped the company's research leadership, and Muse Spark is the first product to carry that organisation's imprint. A strong reception would validate a reorganisation that drew scepticism from rivals and from some inside the company; a muted one would renew questions about the cost of the overhaul.
Meta paired the announcement with a number that underscored the stakes. The company confirmed plans to spend between $115bn and $135bn on AI capital expenditure in 2026, nearly double the prior year, a figure that places it alongside Microsoft, Google and Amazon in an arms race over data centres and computing power that has come to define the industry's economics. The scale of the outlay leaves little room for a flagship model that fails to land.
For Meta, the strategic logic runs through its own products. Better and cheaper models feed directly into the recommendation and advertising systems that generate the company's revenue, and into the assistant features it has been threading through Facebook, Instagram and WhatsApp. The promise of frontier performance at lower cost is, in that sense, less about winning a benchmark leaderboard than about making AI cheap enough to run across platforms used by billions of people every day.