Meta Launches Muse Wonder, Its Biggest Model for Personal and Professional Agents

What happened

On October 12, 2026, Meta introduced Muse Wonder and Muse Wonder Full, calling them its most capable models yet for personal and professional agentic tasks. The release is the next step after Muse Spark on Meta's scaling ladder toward what the company calls personal superintelligence. Both models were trained at Meta's gigawatt-scale Prometheus data center.

This page is the long-form briefing behind the BSH Technologies Instagram carousel.

Two models, two access paths

• Muse Wonder — available today through the Meta Model API and Muse Code.
• Muse Wonder Full — available today for Muse Code subscribers only; no per-token price published.

Meta says Muse Wonder will soon power experiences in the Muse app, rolling out first to subscribers. Inputs can mix text, image, video and audio across a 1 million-token context; answers come back as text.

Active Memory

Meta's name for reliable cross-conversation memory: the model carries important context about you and actively retrieves the details that matter for the current task. On Meta's own MemoryBench (long multi-turn chats of 64K–100K tokens with distractors and conflicting updates), Muse Wonder scores 85.6 and Wonder Full 85.5, ahead of Claude Opus 5.5 at 81.5.

Active Focus

Active Focus is how Muse Wonder keeps working toward a hard goal: trying alternatives, recovering from blockers, asking for clarification, and checking in before consequential actions. Meta evaluates this on FocusBench and sustained software work on SWE-Marathon.

Where Wonder leads (and where it does not)

On Meta's 14-row evaluation table (max effort; Meta notes competitor runs were best-effort), the Wonder family leads 7 rows, Claude Opus 5.5 leads 6, and GPT-6 Astra leads 1. The split matches the product pitch:

• Wonder leads long-horizon work: SWE-Marathon (Wonder Full 57.5% at xhigh vs Opus 5.5 46.3%), DeepSWE 1.1 (79.0 vs 74.2), MemoryBench, AutomationBench-AA, Harvey LAB-AA (Wonder Full 13.1% vs Astra 8.6%).
• Opus 5.5 still leads Terminal-Bench, OSWorld, ProgramBench, HealthBench Pro, and the Artificial Analysis Intelligence Index (58 vs 56 for Wonder Full).

As SiliconANGLE notes, the SWE-Marathon figures are not a fully controlled comparison — Codex and Claude Code produced some rival scores, and Wonder Full ran at an "xhigh" reasoning setting for that test. Most of the winning column is Wonder Full (subscription-only); plain Muse Wonder beats Opus 5.5 on 5 of 14 rows.

Priced like Opus 5.5

Per independent reads of Meta's Model API price list on launch day (CellCog):

• $4.00 per million input tokens
• $0.20 cached input
• $20.00 per million output tokens
• 1,048,576 token context

That matches Claude Opus 5.5's sticker price. Muse Spark 1.3 listed at $1.25 / $4.25, so Wonder is a clear step up the ladder. Prompt caching on Wonder is automatic (no flag); the optional 24-hour retention setting is a hint, not a guarantee.

Safety note

Meta calls Muse Wonder its most aligned model to date and promises a full safety report soon. The model card says Meta does not rule out Wonder Full reaching a high-risk threshold for chemical and biological threats without mitigations; with safeguards, it rates "moderate or lower" across Meta's Superintelligence Scaling Framework domains. On an agentic misalignment test, Wonder Full's harmful-action rate was reported at 0%.

Why it matters

• Personal agents need memory and persistence — Active Memory + Active Focus are Meta's product answer to agents that forget and agents that give up.
• The price signal — Meta set Wonder at Opus parity, not under it, which frames this as a frontier peer rather than a discount challenger.
• Coding is split — Wonder Full looks strong on long-running SWE work; Opus still wins terminal, computer-use and from-scratch program rebuilds on Meta's own chart.

What teams should do now

• Try Wonder on the Meta Model API for long-horizon coding and memory-heavy agent workflows; keep Opus for terminal and computer-use tasks until independent runs land.
• Do not treat Meta's table as gospel — wait for Artificial Analysis and Vals independent listings beyond Meta's launch chart.
• Watch the Muse app rollout and any per-token price for Wonder Full.

Primary sources

• Meta: Muse Wonder — A Frontier Model For Personal & Professional AI Agents (Oct 12, 2026)
• SiliconANGLE: Meta releases Muse Wonder AI models (Oct 12, 2026)
• CellCog: Muse Wonder price and benchmarks (Oct 12, 2026)

How BSH can help

At BSH Technologies we help teams pick and wire frontier models into real agent stacks: routing between providers, evaluating memory and long-horizon coding on your own tasks, and shipping production guardrails. If you want Muse Wonder (or Opus) stress-tested on your workflows, our Thrissur engineers can run that evaluation with you.