Meta drops Muse Wonder to supercharge AI agents

Meta just stopped playing the open-source nice guy and went straight for the frontier throat.

The company unveiled Muse Wonder and Muse Wonder Full, its most capable AI models yet. Built after a 14-month internal overhaul, both models are completely closed — marking a sharp pivot away from the Llama playbook.

Meta trained the duo at its gigawatt-scale Prometheus data center in Ohio. And the target is explicit: personal and professional AI agents that handle messy, multi-step work.

Memory that lasts, focus that refuses to quit

Instead of just chasing raw reasoning scores, Meta built Wonder around two distinct mechanisms: Active Memory and Active Focus.

Active Memory solves the agent amnesia problem. Meta says the model actively tracks relevant context across separate chats over days, weeks, and months, pulling up old preferences right when a task demands them. In Meta's self-reported MemoryBench evaluations, it outranked both GPT-6 Astra and Opus 5.5.

Active Focus handles marathon execution. If a tool fails or a blocker appears, the agent tries alternative paths, asks for clarification, and pauses for permission before taking big actions.

Early demos show it coordinating browser workflows and local apps, from designing running routines to pulling photos from Instagram to order merchandise.

Blazing speed with a subscription catch

Third-party benchmark scores show Meta legitimately reaching the top tier, but with caveats.

On Artificial Analysis's Intelligence Index, Wonder Full scored 56, matching Sonnet 5.5, while standard Wonder hit 52. Early testing highlights blisteringly fast inference speeds and low hallucination rates, mainly because the model admits when it does not know an answer.

Yet on public leaderboards like Live Bench, standard Wonder landed outside the top ten. And in task-based benchmarks, it ranked fifth on the Vals Index.

Meta is also keeping the best gear behind a paywall. Standard Wonder is on the API for $4 per million input tokens and $20 for output. But the flagship Wonder Full model is exclusively reserved for Muse Code subscribers paying $50 or $200 a month.

Why it matters

We think this launch signals the end of Meta's role as a purely altruistic provider of open-weight frontier models. Meta wants its own high-margin agent platform, and it is keeping the crown jewels inside its walls.

For developers, standard Wonder poses an awkward value question. Artificial Analysis measured its cost at $4.52 per task, which is pricier than faster, higher-scoring alternatives like Astra. If you want the real frontier capability, you have to buy a Muse Code seat.

For OpenAI and Anthropic, however, Meta's sheer brute force is terrifying. Meta burned over $31 billion in capital spending last quarter alone and has a family of apps touching 3.6 billion daily users.

With Muse already passing 3.4 million downloads, Meta has a direct pipeline to funnel frontier intelligence into millions of workflows. If Active Memory genuinely holds up in daily enterprise use, Meta might finally turn its massive infra spend into an unassailable agent ecosystem.