Independent Artificial Analysis benchmarks put Meta's Muse Wonder Full within two points of Claude Opus 5.5

2026-10-12

Artificial Analysis's Intelligence Index v4.3 scored Muse Wonder Full (max reasoning) at 56 and Muse Wonder (max) at 52, up from the prior Muse Spark 1.3 (max) score of 48 — a gain reported roughly six weeks after Spark 1.3 shipped. Claude Opus 5.5 leads the index at 58. On Terminal-Bench 4.0, Wonder Full's score rose from 33% to 57%; GDPval-AA v2.1 Elo rose from 1681 to 1857; AA-Briefcase Elo rose from 1581 to 1729. Wonder Full ranked top three on 6 of Artificial Analysis's 10 evaluations, including #1 on GDP.pdf and CritPt, #2 on GDPval, and #3 on AA-Briefcase, AutomationBench-AA, and SciCode. Reported via Latent Space's AINews digest of Artificial Analysis's published results.

Significance 4: New independent (Artificial Analysis) scores on tracked reasoning and agentic axes show Meta closing most of the gap to frontier within six weeks of its prior release, a material capability move even though it does not top the index outright.

technology business Benchmark result Lab strategy

Implications · machine-drafted, not owner judgment

Meta closing from 48 to 52–56 on Artificial Analysis's index in six weeks, with the largest gains on agentic evals (Terminal-Bench, GDPval-AA), suggests Meta Superintelligence Labs' 14-month build-out is translating into measured capability gains rather than only marketing, though this is a single independent measurement firm's reading and not yet cross-checked by a second source. It raises the bar for scrutiny of Meta's own agentic product claims (Active Memory, Active Focus) inside Muse, especially given the already-noted absence of a detailed system card at launch.

Watch for

A second independent benchmark run (e.g. SWE-bench Verified or LMArena) placing Wonder Full within a few points of Opus 5.5
Whether Meta publishes a full system card now that independent numbers invite closer scrutiny

Sources

curator [AINews] Meta Muse Wonder & Wonder Full: MSL's Claude 5.5-tier Frontier Model retrieved 2026-10-13

Affected axes

Hard reasoning · Agentic / tool use · Coding

← Back to feed