Meta Muse Wonder and Muse Wonder Full: A Wall Between Two Models and Three Months of Memory

Forum topic · QianXun · 2026-10-15

Summary

On October 12, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, its first frontier-level models since the 2025 AI reorganization under Alexandr Wang. Trained at the gigawatt-scale Prometheus data center in Ohio, both models accept text, image, video, and audio input with a 1-million-token context window and text-only output. Two named capabilities headline the launch: Active Memory, which retains cross-session context, and Active Focus, which pursues long-horizon goals with checkpointing before irreversible actions. Benchmarks split sharply: Muse Wonder Full leads on long-horizon tasks like SWE-Marathon (57.5% vs Opus 5.5's 46.3%) and MemoryBench (85.5%), but trails on knowledge-heavy benchmarks (30% vs 46%) and Terminal-Bench. Artificial Analysis ranks it at 56 on Intelligence Index v4.3, two points behind Claude Opus 5.5 — frontier-adjacent, not dominant. Notably, the stronger Full model is subscription-only ($50–$200/month Muse Code), while Wonder sells via API ($4/$20 per million tokens), a deliberate wall testing whether users will pay recurring fees for agent capability.

On October 12, Meta Superintelligence Labs unveiled Muse Wonder and Muse Wonder Full, the company's first frontier-class models since its 2025 AI reorganization under Alexandr Wang. Training ran at the Prometheus data center in Ohio, with a 1-million-token context window and multimodal input (text, image, video, audio) yielding text-only output. Insurance claims processing and coding were cited as flagship use cases.

Key points

Two named behaviors define the models: *Active Memory* retains important cross-session context and proactively recalls it; *Active Focus* sustains work toward difficult goals, trying alternatives and confirming with users before critical operations.
A wall separates the two tiers: Muse Wonder is sold via the Meta Model API ($4/M input, $20/M output, $0.20 cached input), while the stronger Wonder Full is available only through Muse Code subscriptions ($50 or $200/month) — CNBC framed this as Meta testing paid agents. No parameters disclosed, no open weights, though Zuckerberg has said open-weight Muse releases are planned.
Independent vs. self-reported scores diverge: Artificial Analysis Intelligence Index v4.3 gives Wonder Full 56 and Wonder 52, versus Claude Opus 5.5 at 58 and Sonnet 5.5 at 56 — within normal jitter. Meta's own table, run at maximum reasoning with best-effort competitor testing, shows larger leads.

Meta self-reported benchmarks (Wonder Full)

| Benchmark | Wonder Full | Best rival | Delta | |---|---|---|---| | SWE-Marathon (long-horizon coding) | 57.5% | Opus 5.5 46.3% | +11.2 pts | | DeepSWE v1.1 | 79.0% | Opus 5.5 74.2% | +4.8 pts | | Harvey LAB-AA (legal agent) | 13.1% | GPT-6 Astra 8.6% | +4.5 pts | | MemoryBench | 85.5% | 81.5% | +4.0 pts | | Terminal-Bench 4.0-AA | 57.1% | Opus 5.5 59.6% | −2.5 pts | | HealthBench Pro | 62.3% | Opus 5.5 65.6% | −3.3 pts | | AA-Omniscience (knowledge) | 30% | Opus 5.5 46% | −16 pts |

The pattern: Meta leads on long-horizon, memory-dependent agentic tasks and trails on single-turn knowledge and precision tasks. A 16-point knowledge gap is not noise. Wonder Full also posted a 23% hallucination rate, among the lowest tested by Artificial Analysis.

Cost and caveats

Artificial Analysis measured Wonder at $4.52 per index task versus Opus 5.5's $5.98, GPT-6 Astra's $3.26, and Gemini 4 Argon's $1.99 — over five times the $0.72 of GPT-6.1 Sol. "Cheap" is not part of this launch.

VentureBeat cautioned that Meta's comparison table mixes internal tests, vendor-supplied, and third-party numbers, sometimes with different tool scaffolding; Meta's model card admits SWE-Marathon comparisons were not controlled. Positively, MemoryBench lives on arXiv, Artificial Analysis independently reproduced the index scores the next day, and the safety disclosure is unusually direct: Meta flagged possible unmitigated high-risk chemistry/bio thresholds, prompting safety post-training and system-level classifiers. CyberGym: 88.4%; ExploitBench: 58.9%; of 92 flagged ExploitGym attempts, only 21 actually exploited target vulnerabilities.

Outlook

Six weeks after Muse Spark 1.3 scored 48 on the index, Wonder Full hits 56 — a rapid rise, but the top seat remains unclaimed. The bet is that memory becomes a first-class citizen of agents, priced behind a monthly fee; whether that bet pays off will show in subscription renewal rates, not benchmarks.

References

Meta Superintelligence Labs, *Muse Wonder release materials and model card*, 2026-10-12. https://ai.meta.com/blog/
Artificial Analysis, *Intelligence Index v4.3 and cost pages*, 2026-10-13. https://artificialanalysis.ai/
Traictory, *Meta's Muse Wonder is a frontier model family now*, 2026-10-14. https://traictory.com/news/2026-10-14-meta-muse-wonder-frontier-models
InstaDataNews, *Meta Muse Wonder Is Its First Frontier Model, and It's Closed*, 2026-10-15. https://instadatanews.com/articles/2026-10-15-meta-muse-wonder-first-frontier-model.html
VentureBeat / TechCrunch / CNBC coverage, 2026-10-15. https://www.163.com/dy/article/L9A32OTF05118UGF.html