Meta unveils Muse Wonder and Muse Wonder Full as its most capable AI models to date, designed to boost personal AI assistants and coding tools while signaling a broader push into agentic capabilities.

Analysts estimate Wonder costs about $4.52 per Intelligence Index task, making it pricier than Spark 1.3 yet cheaper than some rivals, with a mix of input/output token dynamics that favor input tokens over its predecessor.

A standout feature is active or long-term memory, aiming to recall user preferences over days, weeks, or months, supported by MemoryBench as evidence of ongoing continual learning progress.

Industry analysis frames the rollout in terms of economics and strategy for advertisers, noting that ad revenue funds infrastructure and future pricing may hinge on edge versus central data-center compute.

Independent verification and a full safety report are still pending, with limited external testing on long-running tasks, privacy behavior, and benchmarks.

Meta reported its Q2 results with earnings per share of $6.18 on revenue of $60.80 billion, beating neither consensus nor expectations and signaling mixed near-term performance while year-over-year revenue rose about 28%.

Industry observers say momentum is shifting from chatbots to agentic systems, but enterprise hesitancy over trust, data privacy, and integration remains a hurdle.

Some sections of the article are inaccessible due to page design, hindering access to specific dates, figures, or quotes from the main text.

IMF warns that revenue commitments behind AI deals could create wrong-way risk if expected revenues fail to materialize, posing broader financial risk.

DRAM requirements could strain memory supply, potentially representing about 5% of the world’s annual DRAM output if scaled to large platforms.

Market sentiment remains mixed, with analysts weighing potential upside against execution risk in the AI tooling and memory supply chain.

Internal benchmarks are impressive, but real-world adoption will depend on developers’ experiences with data access and handling error conditions.