Meta keeps its best model closed: Muse Wonder scores 56 vs Opus 5.5's 58 at $4/$20

Muse Wonder is Meta's proprietary frontier tier. On the Artificial Analysis Index it scores 56, two points behind Claude Opus 5.5 at 58, and it costs $4 per million input tokens and $20 per million output tokens. Meta's pitch rests on three capabilities: deep reasoning, dynamic memory, and native execution across multiple tools.

Bloomberg's framing is that Meta is releasing an open model while keeping its best technology closed. That marks a strategic break from the Llama-era promise that Meta's frontier work would be open. Yahoo Finance reports that Meta's CFO warned the stakes are 'really bad' if the AI bet flops, so the company now needs revenue from its models, not just goodwill.

HN commenters see price as the real story. One wrote that 'it'll become a commodity and the price will be driven to the ground.' Another: 'If this model is not benchmaxxed, I fear it's over for anthropic, openai etc.' That fits the broader 'frontier parity' mood: when the top labs cluster within a few index points, cost and tooling decide who wins.

The caveat is the one in that HN comment. A single aggregate index can hide weaknesses, and real-world agentic evaluations are still to come.