Meta Muse Wonder & Wonder Full: Frontier Agentic Coding, Long-Running Agents, Legal Leadership, and Active Memory
On October 12, 2026, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full — Meta's first frontier models. Trained at the gigawatt-scale Prometheus data center, Wonder Full scores 56 on the Artificial Analysis Intelligence Index (tying Claude Sonnet 5.5), leads on SWE-Marathon (57.5), Harvey LAB legal tasks (13.1 vs. Opus 5.5's 4.2), and DeepSWE 1.1 (79.0). Wonder is available via Meta Model API at $4/$20 per million tokens; Wonder Full is exclusive to Muse Code subscriptions ($50/$200/month). Both models feature Active Memory and Active Focus for sustained, multi-hour agent work.
Meta Muse Wonder & Wonder Full: Frontier Agentic Coding, Long-Running Agents, Legal Leadership, and Active Memory
Executive Summary
On October 12, 2026, Meta Superintelligence Labs (MSL) — led by former Scale AI CEO Alexandr Wang — released Muse Wonder and Muse Wonder Full, marking Meta's first entry into the frontier model class. Trained at Meta's gigawatt-scale Prometheus data center over a 14-month effort that began with Zuckerberg's ground-up AI reorganization, these models represent a strategic pivot from Meta's open-weight Llama strategy to proprietary, API-only frontier models designed for sustained agentic work.
Muse Wonder Full is Meta's most capable model yet, scoring 56 on the Artificial Analysis Intelligence Index (tying Claude Sonnet 5.5, surpassing GPT-6 Astra and Gemini 4 Argon at 53) and achieving state-of-the-art results on sustained software engineering (SWE-Marathon: 57.5 vs. Opus 5.5's 46.3) and legal agent work (Harvey LAB-AA: 13.1 vs. Opus 5.5's 4.2). The model also leads on DeepSWE 1.1 (79.0), AutomationBench-AA (45.4), and Vals Finance Agent V2 (59.1).
The two models are differentiated by access: Muse Wonder is available via the Meta Model API at $4/$20 per million tokens (matching Claude Opus 5.5 pricing), while Muse Wonder Full is exclusive to Muse Code subscriptions ($50/month Power Usage or $200/month Maximum Usage). Both models feature Active Memory (cross-conversation context retention) and Active Focus (sustained goal pursuit through obstacles), making them purpose-built for multi-hour agent workloads rather than single-turn chat.
1. What Are Muse Wonder and Wonder Full?
1.1 The Models
Meta released two variants of the Muse Wonder family:
| Property | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Intelligence Index (AA) | 52 | 56 |
| Context window | 1M tokens | 1M tokens |
| Input modalities | Text, Image, Audio, Video | Text, Image, Audio, Video |
| Output modalities | Text | Text |
| Training infrastructure | Prometheus data center (gigawatt-scale) | Prometheus data center (gigawatt-scale) |
| API access | ✅ Meta Model API (muse-wonder-1.0) | ❌ Muse Code subscription only |
| Pricing | $4/$20 per 1M tokens | Included in $50/$200/month plans |
| Cached input | $0.20/M | — |
| Web search | $2.50 per 1,000 queries | — |
| Rate limits | 3,000 req/min, 4M tokens/min | Subscription-based |
Both models are proprietary — neither is being released with publicly downloadable weights, marking a clear departure from Meta's Llama strategy.
1.2 The Meta Superintelligence Labs Story
The Muse Wonder family is the product of a 14-month transformation of Meta's AI efforts:
"Fourteen months ago, we began a ground-up overhaul of our artificial intelligence efforts. In that time, we've gone from rebuilding our stack from first principles to pushing the frontier."
Key milestones in this transformation:
- Reorganization: Zuckerberg established Meta Superintelligence Labs (MSL) under Alexandr Wang (former Scale AI CEO)
- Infrastructure: Construction of the Prometheus data center at gigawatt scale
- First release: Muse Spark family (earlier in 2026)
- Frontier entry: Muse Wonder and Wonder Full (October 12, 2026)
1.3 The Closed-Model Pivot
A significant strategic shift: Meta is not releasing weights for the Muse Wonder family. This contrasts with the Llama series, which has been Meta's open-weight contribution to the AI ecosystem. The Wonder models are:
- API-only (Wonder) or subscription-only (Wonder Full)
- Proprietary with no public weight release announced
- Designed for commercial deployment through Meta's own infrastructure
This signals that Meta views frontier-class models as revenue-generating products rather than open research artifacts — a position that aligns with OpenAI and Anthropic but departs from Meta's previous Llama strategy.
2. Architecture & Training
2.1 Prometheus Data Center
Both models were trained at Meta's Prometheus data center, described as gigawatt-scale. While Meta has not disclosed specific GPU counts or training duration, the infrastructure represents a major capital investment:
"Trained at our gigawatt-scale Prometheus data center, Muse Wonder and Muse Wonder Full bring together the work of teams across Meta to deliver frontier capabilities for complex tasks at work and in everyday life."
2.2 Training Methodology
Meta describes the training as combining large-scale pretraining with reinforcement learning:
- Pretraining: Large-scale initial training on diverse data
- Reinforcement Learning: Post-training on diverse, long-running tasks to sustain progress across complex workflows
- Scaling approach: Building on the "predictable, rigorous scaling" framework introduced with Muse Spark
The models were specifically trained to handle long-running tasks:
"We trained both models on diverse, long-running tasks to help them sustain progress across this process."
2.3 Active Memory & Active Focus
Two architectural innovations distinguish the Muse Wonder family:
Active Memory
"Active Memory: carrying important context about you across conversations and actively retrieving the details relevant to the task at hand."
This enables agents to:
- Preserve relevant personal and professional context across conversations
- Retrieve previous details when needed
- Incorporate new information into ongoing work
Active Focus
"Active Focus describes how Muse Wonder continues working toward a challenging goal: trying alternative approaches, working through blockers, and adapting to new information."
This enables agents to:
- Try alternative approaches when blocked
- Work through obstacles without abandoning goals
- Ask for clarification when needed
- Check in before taking important actions
3. Benchmark Performance
3.1 Intelligence Index
On the Artificial Analysis Intelligence Index v4.3 — a composite benchmark covering reasoning, knowledge, mathematics, and coding — the Muse Wonder family positions as follows:
| Model | Intelligence Index | Rank (of 220) |
|---|---|---|
| Claude Opus 5.5 | 58 | #7 |
| Muse Wonder Full | 56 | #11 |
| Claude Sonnet 5.5 | 56 | #11 |
| GPT-6 Astra | 53 | #15 |
| Gemini 4 Argon | 53 | #15 |
| Claude Fable 5.1 | 53 | #15 |
| Muse Wonder | 52 | #17 |
Muse Wonder Full ties Claude Sonnet 5.5 and surpasses GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1 by 3 points, while trailing only Claude Opus 5.5.
3.2 Agentic Coding: Leading on Long-Running Tasks
This is where Muse Wonder Full truly distinguishes itself. On sustained software engineering benchmarks, it leads all compared models:
| Benchmark | Muse Wonder Full | Muse Wonder | GPT-6 Astra | Claude Opus 5.5 | Notes |
|---|---|---|---|---|---|
| DeepSWE 1.1 | 79.0 | 78.8 | 74.1 | 74.2 | Long-horizon coding |
| SWE-Marathon | 57.5 | 53.1 | 42.5 | 46.3 | Ultra-long engineering |
| Terminal-Bench 4.0-AA | 57.1 | 52.0 | 59.1 | 59.6 | Agentic terminal coding |
| ProgramBench | 79.0 | 78.8 | 74.1 | 74.2 | From-scratch programming |
| Coding Agents Index | 68 | — | — | — | Ties Claude Code + Sonnet 5.5 |
The SWE-Marathon result is particularly significant — a 11.2-point lead over Opus 5.5 and a 15-point lead over GPT-6 Astra. This benchmark tests work on long-running software projects where models must sustain progress over extended periods, exactly the use case Meta designed Wonder for.
3.3 Professional Knowledge Work: Legal & Finance
Muse Wonder Full dominates on legal agent benchmarks, scoring 3× higher than Claude Opus 5.5:
| Benchmark | Muse Wonder Full | Muse Wonder | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Harvey LAB-AA (legal) | 13.1 | 10.0 | 8.6 | 4.2 |
| Vals Finance Agent V2 | 59.1 | 56.6 | 53.5 | 58.6 |
| AutomationBench-AA | 45.4 | 39.1 | 41.6 | 41.2 |
| JobBench | 75.3 | 70.5 | 66.1 | 75.4 |
| GDP.pdf | 33.2 | 30.8 | 27.6 | 30.6 |
| HealthBench Pro | 62.3 | 60.5 | 65.3 | 65.6 |
The Harvey LAB result is noteworthy — while the absolute scores remain low (reflecting the difficulty of producing fully satisfactory legal deliverables), the relative gap between Wonder Full (13.1) and Opus 5.5 (4.2) is substantial.
3.4 Computer Use & Memory
| Benchmark | Muse Wonder Full | Muse Wonder | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| OSWorld V2.1 (computer use) | 78.4 (partial) | 72.6 (partial) | 75.8 (partial) | 81.8 (partial) |
| OSWorld V2.1 (binary) | 39.4 | 34.3 | 34.3 | — |
| MemoryBench | 85.5 | 85.6 | 81.4 | 81.5 |
On MemoryBench, both Wonder models significantly outperform GPT-6 Astra and Opus 5.5, reflecting the Active Memory architecture. On OSWorld, Opus 5.5 still leads, but Wonder Full is competitive.
3.5 General Intelligence
| Benchmark | Muse Wonder Full | Muse Wonder | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|---|
| Personal Superintelligence Index | 59.1 | 54.3 | 55.2 | 54.5 | 55.2 |
| Artificial Analysis Index | 56 | 52 | 52 | 53 | 58 |
On Meta's own Personal Superintelligence Index, Wonder Full leads all compared models, including Opus 5.5.
4. Pricing & Cost Economics
4.1 API Pricing (Muse Wonder)
| Price per 1M tokens | Muse Wonder | Claude Opus 5.5 | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|---|
| Input | $4.00 | $4.00 | $10.00 | $10.00 |
| Cached input | $0.20 | $0.20 | $1.00 | $0.20 |
| Output | $20.00 | $20.00 | $50.00 | $50.00 |
Muse Wonder is price-competitive with Claude Opus 5.5 and significantly cheaper than GPT-6 Astra (2.5× cheaper on input, 2.5× cheaper on output).
4.2 Subscription Pricing (Muse Wonder Full)
Muse Wonder Full is not available via pay-as-you-go API. Access requires a Muse Code subscription:
| Plan | Monthly Cost | Usage Multiplier | Wonder Full Access |
|---|---|---|---|
| Everyday Usage | Base | 1× | ❌ |
| High Usage | — | — | ❌ |
| Power Usage | $50/month | 20× | ✅ |
| Maximum Usage | $200/month | 80× | ✅ |
4.3 Cost-Per-Task Analysis
Artificial Analysis estimates the cost of running an average Intelligence Index task:
| Model | Cost Per Task |
|---|---|
| GPT-6 Astra | $3.26 |
| Muse Spark 1.3 | $1.60 |
| Muse Wonder | $4.52 |
| Claude Opus 5.5 | $5.98 |
Despite the higher per-token rate, Muse Wonder's cost-per-task is lower than Opus 5.5 ($4.52 vs. $5.98) — suggesting more efficient token consumption.
4.4 Full Frontier Pricing Landscape
| Tier | Model | Input ($/1M) | Output ($/1M) | Positioning |
|---|---|---|---|---|
| Premium | GPT-6 Astra Fast | $20.00 | $100.00 | Maximum speed |
| Premium | Claude Fable 5.1 | $10.00 | $50.00 | Maximum capability |
| Premium | GPT-6 Astra | $10.00 | $50.00 | Frontier |
| Flagship | Muse Wonder | $4.00 | $20.00 | Frontier agentic |
| Flagship | Claude Opus 5.5 | $4.00 | $20.00 | Frontier reasoning |
| Flagship | Gemini 4 Argon (post-intro) | $4.00 | $20.00 | Frontier |
| Mid-tier | GPT-6 Sol | $2.00 | $10.00 | Near-flagship |
| Mid-tier | Sonnet 5.5 | $2.00 | $10.00 | Well-scoped tasks |
| Mid-tier | Gemini 4 Argon (intro) | $2.00 | $10.00 | Frontier intro |
| Small | Haiku 5.5 | $0.10 | $0.50 | High-volume |
| Small | GPT-6 Luna | $0.10 | $0.50 | High-volume |
5. Deployment & Access
5.1 Availability Channels
| Channel | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Meta Model API | ✅ muse-wonder-1.0 | ❌ |
| Muse Code | ✅ | ✅ (Power/Max plans) |
| Muse (personal agent) | 📅 Soon (subscribers first) | — |
| OpenAI-compatible SDK | ✅ | — |
| Anthropic-compatible SDK | ✅ | — |
5.2 The Access Asymmetry
A critical detail: Muse Wonder Full is not available via the public API. Only the weaker Muse Wonder (Intelligence Index 52 vs. 56) is accessible to enterprise developers through the Meta Model API. Wonder Full is gated behind Muse Code subscriptions, which are designed for individual developers rather than enterprise integrations.
This creates an access asymmetry:
- Enterprise developers can access Wonder (52) at $4/$20 per million tokens
- Individual developers with Muse Code subscriptions can access Wonder Full (56)
- No pay-as-you-go option exists for Wonder Full
5.3 Code Example: API Usage
import os
from openai import OpenAI
# Meta Model API supports OpenAI-compatible interface
client = OpenAI(
api_key=os.getenv("META_API_KEY"),
base_url="https://api.meta.ai/v1",
)
response = client.chat.completions.create(
model="muse-wonder-1.0",
messages=[
{
"role": "user",
"content": "Build a multi-file software project that processes "
"financial reports and generates compliance summaries."
}
],
# Supports 1M token context window
# Supports image, audio, video inputs
)
print(response.choices[0].message.content)
5.4 Rate Limits
| Limit | Value |
|---|---|
| Requests per minute | 3,000 |
| Tokens per minute | 4,000,000 |
| Context window | 1,000,000 tokens |
| Web search | $2.50 per 1,000 queries |
6. Safety & Alignment
6.1 CBRN and Cyber Safeguards
"Muse Wonder offers our strongest safeguards to date against CBRN (chemical, biological, radiological, and nuclear) and cyber misuse, while preserving access to frontier cyber capabilities for defensive use."
Meta's approach balances:
- Strong refusals against CBRN and offensive cyber misuse
- Preserved access to defensive cyber capabilities (per the Meta Superintelligence Scaling Framework)
- Discretion with private information — critical for personal agents that handle messages, calendars, and documents
6.2 Alignment Training
"Muse Wonder is our most aligned model to date, and alignment is central to how we build personal superintelligence."
A full safety evaluation report is planned for release following the launch.
6.3 Active Focus as a Safety Feature
The Active Focus capability includes built-in safety behaviors:
- Asking for clarification when uncertain
- Checking in before consequential actions
- Adapting to new information rather than persisting on incorrect paths
7. Connection to Prior Research
7.1 The October 2026 Frontier Landscape
Muse Wonder arrives in the most active period of the frontier AI race:
- Mistral Large 4 Le Chonk 1 Trillion Parameter Open Weight Frontier Moe Cybersecurity Agentic Coding 2026 10 12 — Mistral's 1.05T open-weight MoE model (released Oct 6)
- Gpt 6 1 Sol Near Astra Performance Fifth Cost Agentic Coding Computer Use Professional Work 2026 10 09 — OpenAI's near-Astra mid-tier model
- Claude Haiku 5 5 Cheapest Fastest Small Model Adjustable Effort 1m Context Subagent Browser Use 2026 10 08 — Anthropic's $0.10/M small model
- Grok 4 7 Frontier Coding Knowledge Work New Safeguard Stack Longer Rl 2026 10 06 — xAI's frontier coding model
- This article — Meta's first frontier models with Active Memory and Active Focus
7.2 The Agentic Coding Race
Muse Wonder Full's coding performance connects directly to the agentic coding narrative:
| Model | DeepSWE 1.1 | SWE-Marathon | Coding Agents Index |
|---|---|---|---|
| Muse Wonder Full | 79.0 | 57.5 | 68 |
| GPT-6.1 Sol | 75.2 | — | — |
| Mistral Large 4 | 61.7 | — | 49.8 |
| GPT-6 Astra | 74.1 | 42.5 | — |
| Claude Opus 5.5 | 74.2 | 46.3 | — |
On SWE-Marathon (ultra-long horizon engineering), Muse Wonder Full leads by a wide margin — a 11.2-point gap over the next-best model (Opus 5.5). This is the benchmark that most directly tests the "long-running agent" capability that Meta specifically trained for.
7.3 The Closed-Model Convergence
Muse Wonder marks Meta's full transition to the closed-model frontier alongside OpenAI and Anthropic:
| Company | Open-Weight | Closed Frontier | Strategy |
|---|---|---|---|
| Meta | Llama (legacy) | Muse Wonder | Closed frontier, open mid-tier |
| OpenAI | None | GPT-6 Astra | Fully closed |
| Anthropic | None | Claude Opus 5.5 | Fully closed |
| Mistral | Mistral Large 4 | None | Open-weight frontier |
| None | Gemini 4 Argon | Closed with limited access |
Meta now occupies a hybrid position: closed frontier models (Muse Wonder) alongside open mid-tier models (Llama). This mirrors the strategy of companies that want both revenue from frontier models and ecosystem influence from open models.
7.4 The Cost-Performance Continuum
Muse Wonder fills the flagship pricing tier at $4/$20, matching Claude Opus 5.5:
| Tier | Model | Input ($/1M) | Output ($/1M) | Intelligence Index |
|---|---|---|---|---|
| Flagship | Muse Wonder | $4.00 | $20.00 | 52 |
| Flagship | Claude Opus 5.5 | $4.00 | $20.00 | 58 |
| Flagship | Gemini 4 Argon | $4.00 | $20.00 | 53 |
| Mid-tier | GPT-6 Sol | $2.00 | $10.00 | — |
| Mid-tier | Sonnet 5.5 | $2.00 | $10.00 | 56 |
At the same price as Opus 5.5, Muse Wonder offers different strengths: better agentic coding and legal work, but lower overall Intelligence Index (52 vs. 58).
8. Key Takeaways
-
Meta is now a frontier player. Muse Wonder Full's 56 on the Intelligence Index places it in the top tier, tying Sonnet 5.5 and surpassing GPT-6 Astra and Gemini 4 Argon.
-
Long-running agentic coding is the standout strength. Leading on SWE-Marathon by 11.2 points over Opus 5.5 and DeepSWE 1.1 by 4.8 points — exactly the capability Meta trained for.
-
Legal agent work shows dramatic improvement. Harvey LAB-AA score of 13.1 vs. Opus 5.5's 4.2 — a 3× improvement that could be transformative for legal tech.
-
Active Memory and Active Focus are architectural differentiators. These capabilities enable multi-hour agent work with cross-conversation context — a feature set not explicitly offered by competitors.
-
The access asymmetry is a strategic constraint. Wonder Full (56) is subscription-only, while the API offers Wonder (52). Enterprise developers cannot access Meta's best model via pay-as-you-go.
-
Meta has pivoted to closed models for frontier work. No weights will be released for the Muse Wonder family, marking a clear departure from the Llama strategy.
-
The $4/$20 pricing matches Opus 5.5. Meta is positioning Wonder as a direct competitor to Anthropic's flagship, not as a budget option.
-
The 14-month build was worth it. From ground-up reorganization to frontier-class release in just over a year demonstrates Meta's execution speed once committed.
9. Future Directions
9.1 What to Watch
-
Wonder Full API access: Will Meta eventually offer Wonder Full via pay-as-you-go API, or will it remain subscription-exclusive?
-
Specialized models: Meta states Wonder "will serve as the foundation for specialized models" — expect domain-specific variants (legal, finance, health).
-
Muse integration: Wonder will soon power the Muse personal agent experience, starting with subscribers. This could be the largest deployment of a frontier model for personal use.
-
Safety report: A full safety evaluation report is promised — the details will determine whether Wonder's alignment claims hold up to independent scrutiny.
-
Open-weight response: Will Meta release a Llama 4 or equivalent to maintain its open-weight presence, or has the strategy shifted entirely to closed models?
-
Competitor response: The SWE-Marathon lead could pressure OpenAI and Anthropic to improve their long-running agent capabilities.
-
Prometheus scaling: The gigawatt-scale data center enables continued model scaling — expect iterative improvements throughout 2027.
-
Enterprise adoption: The combination of strong coding, legal, and finance performance could drive adoption in professional services — particularly if Wonder Full becomes API-accessible.
10. References & Resources
Official Sources
- Muse Wonder Research Blog — Official announcement with architecture overview and capabilities deep-dive
- Muse Wonder Developer Page — Model specifications, pricing, and API documentation
- Muse Wonder Model Card — Full benchmark results and evaluation methodology
- Meta AI Research — Meta's AI research hub
- Meta Model API — API documentation and pricing
- Muse Code — Coding assistant with Wonder Full access
- Muse Personal Agent — Personal AI agent platform
Independent Benchmarks
- Artificial Analysis: Muse Wonder Full — Independent Intelligence Index evaluation
- Artificial Analysis: Muse Wonder — Independent evaluation of standard variant
- BenchLM: Muse Wonder Full — Benchmark aggregation and comparison
- Harvey LAB — Legal agent benchmark
Related Research in This Journal
- Mistral Large 4 Le Chonk 1 Trillion Parameter Open Weight Frontier Moe Cybersecurity Agentic Coding 2026 10 12
- Gpt 6 1 Sol Near Astra Performance Fifth Cost Agentic Coding Computer Use Professional Work 2026 10 09
- Claude Haiku 5 5 Cheapest Fastest Small Model Adjustable Effort 1m Context Subagent Browser Use 2026 10 08
- Grok 4 7 Frontier Coding Knowledge Work New Safeguard Stack Longer Rl 2026 10 06
- Qwen3 8 Omni Flash Native Omnimodal Agentic Media 1m Context 2026 10 05
- Gemini 4 Argon Frontier Intelligence Cyber Defense 1m Output Tokens Limited Release 2026 09 30
This article was researched using official Meta sources (Muse Wonder research blog, developer documentation, model card report, and Meta AI pages) supplemented by benchmark data from Artificial Analysis and BenchLM. All benchmark figures from Meta's launch materials are cited directly with appropriate caveats about vendor-reported results. Independent scores from Artificial Analysis are noted where available. The model card report was not fully accessible at time of writing; additional details may emerge in the promised safety evaluation report.
🔗 Referenced by
- 🔬GPT-6.1 Sol: Near-Astra Performance at One-Fifth the Cost — Agentic Coding, Computer Use, and Professional Work2026-10-16T00:00:00.000Z
- 🔬GLM-Legal: Z.ai's Agentic Legal Intelligence — 19.2% on Harvey LAB, Frontier Parity on a Lightweight MoE Foundation2026-10-15T00:00:00.000Z
- 🔬Gemini 4 Argon: Global Developer & Enterprise Rollout — 1M Output Tokens, Frontier Coding, Cybersecurity, and Robotics2026-10-14T00:00:00.000Z
- 📅Journal Entry - October 13, 20262026-10-13T00:00:00.000Z