Mark Zuckerberg had to commit to the computers before he knew how good the models would be. In a Cheeky Pint interview published October 12, 2026, he and Alexandr Wang describe the infrastructure decisions facing Meta in late 2025. Zuckerberg puts the scale, loosely, at $100 billion to $200 billion. Those numbers describe infrastructure, not the price of hiring Wang.
That is the useful backdrop to Muse Wonder, the model family Meta released on October 12. The new models give Zuckerberg’s bet something firmer than an ambitious story: an independently measured improvement. They do not yet tell us whether the infrastructure will earn its keep.
The people, the model, the machines
Wang joined Meta after its June 2025 investment in Scale AI. In the interview, Zuckerberg explains the leadership choice in terms of strategic clarity, recruiting and adaptability, rather than a record of personally running a frontier-model lab. That is his account of why he chose Wang, not evidence that one executive caused the results now arriving.
The technical plan connects three things that are easy to discuss separately: who builds the model, what the product asks it to do, and how much computing capacity must be ready beforehand. Waiting for a convincing model would make the hardware decision easier. It could also leave the team waiting for the machines when it was ready to scale.
Meta says Wonder was trained at its gigawatt-scale Prometheus data center. Its research announcement describes improvements in both initial training and reinforcement learning, which rewards useful behavior during further training. The stated destination is an agent that remembers relevant context and keeps working toward a goal, rather than simply producing a good answer to one prompt.
The improvement is not just Meta saying so
Artificial Analysis tested the new models. On its Intelligence Index v4.3, Muse Wonder scores 52 and Wonder Full 56, compared with 48 for Muse Spark 1.3. All three figures use maximum reasoning effort. Full improves across all ten evaluations in that index.
The gain also appears in the coding-agent setup. With Muse Code, Wonder scores 63 and Full 68 on the firm’s Coding Agents Index, against 54 for Spark 1.3. These are index scores, not percentages of all jobs an agent can complete. They establish a stronger tested system, not a universal success rate.
That distinction matters without making the progress small. An outside evaluator has found a meaningful step forward where Meta wants to compete. What it cannot isolate is how much came from leadership, training choices, more compute or the software surrounding the model. A benchmark is evidence about the resulting system, not a controlled experiment in management.
Muse is the product. Wonder is not yet everywhere in it.
Meta introduced Muse in September as a personal agent, reachable through an app and WhatsApp, with a browser and virtual computer for tasks. Wonder is a new model family, not the launch of that entire product.
The Wonder release makes the distinction concrete. Wonder is available through Meta’s model API and Muse Code; the more capable Wonder Full is confined to Muse Code. Integration into the personal Muse experience is still announced as coming, starting with subscribers. Meta’s developer documentation labels Wonder a preview.
Meta calls its two agent capabilities Active Memory and Active Focus: retrieving relevant context, and sustaining work while correcting mistakes or asking for help. Those are sensible targets. They are not a guarantee that an agent will remember everything or finish a long assignment without intervention. Meta’s model card itself describes continuing limitations in multistep work.
For a user, the meaningful unit is a finished task. Did the agent retain the right constraint, notice a mistake and stop when it needed a decision? As our look at AI task-duration benchmarks explains, a longer-looking capability can depend heavily on the test and its ruler.
A better model still needs an economic job
Wonder costs $4 per million input tokens and $20 per million output tokens, with cached input priced separately at $0.20, according to Meta’s current product page. Full has no API token price; access is offered through the $50 and $200 monthly Muse Code tiers.
Artificial Analysis calculates $4.52 per Intelligence Index task for Wonder, versus $1.60 for Spark 1.3. That is a cost within this particular benchmark, not the price of a finished customer assignment. It nevertheless prevents an easy shortcut: the performance advance is not automatically a cost reduction.
Infrastructure adds another layer. The right question is not only how powerful a chip or model is, but what the complete system delivers for its cost. Microsoft’s Maia 300 rack comparison shows why the unit of comparison matters.
Zuckerberg also rejects an all-or-nothing reading of the bet. In the interview, he argues that compute could serve other Meta applications or outside users. That gives his investment story more than one possible route to value, though it does not establish what those alternatives will earn.
Muse Wonder makes the technical case for the rebuild stronger. The remaining gap is between a system that scores better and one that completes enough valuable work to justify the machines behind it. Wang’s team has delivered a measurable gain. Zuckerberg’s larger bet still depends on what people can do with it.
