Meta AI Review

Distribution is the whole argument here. Meta AI does not have to be found, downloaded or paid for, because it already sits in the search bar of four apps that most of the planet opens every day, and it answers inside the thread. The standalone app landed on April 29, 2025 at LlamaCon with full duplex voice, image generation and image editing, and it had logged around sixty million installs by April 2026, a number that sounds enormous until it is set beside the billion monthly users Meta counts across the family of apps.

The model underneath has changed more than the product has. Llama 4 powered it through 2025 and into 2026, and in April 2026 Meta Superintelligence Labs put Muse Spark in front of the chatbots on Meta services, with Muse Image and Muse Video following by July. The assistant got better at voice and at pictures. It did not get better at work. There is no file workspace, no serious tool use and no agent loop worth the name, and the repetitive tasks grade records a thing that answers questions rather than clears a queue.

What it is good at is the low stakes question asked in the middle of something else: translate this, summarise that, make a picture of the dog as an astronaut. What it is not good at is anything a paid assistant would be asked to do, and the habit of placing it where nobody asked for it has not won it goodwill. Summarising other people's journalism without linking to it has not either. It costs nothing, which is both the offer and the ceiling.