Meta Introduces Muse Wonder Frontier Models: Active Memory, Adaptive Task Execution, and Agentic Coding to Rival Opus 5.5

Meta has launched the Muse Wonder frontier model series. This release includes two flagship models: Muse Wonder and the more capable Muse Wonder Full. Both models feature a one-million-token reasoning context window, support multimodal inputs across text, images, audio, and video, output text only, and were trained at the gigawatt-scale Prometheus data center.

Key functional features include:

• Active Memory: Retains important context across sessions, retrieves relevant information in new tasks, and identifies factual updates.
• Active Focus: Adapts plans when encountering obstacles, tracks multi-step tasks, asks questions when necessary, and confirms before important actions.
• Domain Expertise: Draws on contracts, financial reports, and health records to analyze problems and draft documents for review.
• Frontier Coding: Tracks requirements, modifies code across files, tests and debugs iteratively, and supports long-running projects.

Meta positions it as a step toward "personal superintelligence": enabling AI to draw on personal context and persistently handle everyday tasks and professional work.

Where can the two versions be used today?

• Now: Muse Wonder is available via the Meta Model API and Muse Code; Full is accessible only to Muse Code subscribers.
• Upcoming: Wonder will be integrated into Muse, with priority access for subscribers, powering capabilities requiring higher intelligence such as generating interactive applications. Meta also plans to develop dedicated models based on it for all Muse users.

What problems do the four capabilities address?

Long-running tasks require linking earlier context, in-flight changes, and final outcomes. Remembering instructions is only the starting point; strategies must adapt along the way, and deliverables must be checked for correctness before handoff.

1. Active Memory: Remembering context and updating with new requirements

Having to re-explain background details for every new task makes an assistant difficult to use. Active Memory retains critical information across sessions, proactively retrieves it when relevant tasks arise, and recognizes updates. For example, when planning a business trip, it carries over your stated travel preferences, but adopts new conditions once your requirements change.

A one-million-token context corresponds to the volume of material processed in a single inference pass. Cross-session memory must additionally carry useful context into future sessions, retrieving it when needed.

Meta's internal MemoryBench uses long conversations of 64K–100K tokens, injects irrelevant chatter, conflicting facts, and subsequent corrections, and scores performance via an LLM judge using weighted criteria to test whether models can retrieve valid details from noisy information and identify factual updates.

Both models score similarly and outperform all listed comparison models; long-conversation memory is a shared strength.

2. Active Focus: Adapting methods when encountering obstacles

Midway through a task, tools may throw errors, data may be missing, or users may change their minds. Active Focus explores alternative approaches around the goal and adjusts subsequent steps based on new information; it clarifies with users when critical conditions are ambiguous and asks for confirmation before significant actions. Plans must update alongside execution results to keep tasks moving forward.

FocusBench is Meta's internal dynamic scenario benchmark covering domains such as housing, travel, IT, and finance, evaluating persistence and adaptation after encountering obstacles. AutomationBench-AA checks final states across simulated business applications, counting a run as successful only if all verification checks pass. The former evaluates adaptability to change, while the latter measures end-to-end procedural correctness.

Full leads the listed comparison models on both benchmarks, outperforming standard Wonder by 9.3 and 6.3 percentage points, respectively; standard Wonder trails Astra and Opus. Compared to their closely matched memory scores, sustained execution represents Full's more distinct improvement. Still, AutomationBench's end-to-end pass rate remains at 45.4%, falling short of half.

Operating real-world software is evaluated by OSWorld 2.1. In an Ubuntu desktop virtual machine, models complete 108 workflows using screenshots, clicks, keystrokes, and scrolls, reporting both partial-credit scores and full completion rates.

Partial-credit scores reward partial progress, whereas the full completion rate only counts fully successful tasks. Full scores 78.4% and 39.4% on the two metrics; Opus and Fable achieve higher partial-credit scores, and Fable also achieves a higher full completion rate.

3. Domain Expertise: Interpreting materials and completing professional tasks

Domain Expertise applies to handling contracts, financial reports, and health records: grounding findings in original documents, then analyzing or drafting deliverables as specified. Because legal, financial, and healthcare tasks have distinct requirements, they are evaluated separately.

Legal: Harvey LAB-AA 1.1. Models generate legal deliverables based on case materials. An evaluation system reviews task requirements and hallucination issues item by item, counting a run as successful only when all criteria are met.

Both models lead the listed comparison models, but Full's pass rate remains at 13.1%, with most deliverables failing to satisfy all criteria simultaneously.

Finance: Vals Finance Agent V2. Models perform analyses by combining financial reports and market materials, with accuracy measured on a private Vals AI test suite.

Full edges out Fable by 0.2 percentage points and lands close to Opus, showing only a narrow advantage. Standard Wonder trails Full, Fable, and Opus.

Health: HealthBench Pro. Without calling external tools, models answer 525 specialized health questions, graded by an LLM against physician-designed criteria with adjustments for answer length.

Both Wonder models score lower than Opus and Astra, trailing in health Q&A.

4. Frontier Coding: Multi-file modifications, iterative testing, and debugging

Software development demands continuously tracking requirements, comprehending codebases, modifying interdependent files, verifying new features and existing behavior with tests, and fixing bugs based on error outputs. Frontier Coding applies memory, domain expertise, and active focus across this entire workflow. Meta trained both models on diverse, long-horizon software engineering tasks to support sustained development ranging from small apps to complex codebases.

DeepSWE 1.1 covers 113 tasks across 91 repositories and 5 languages, using functional and regression tests to determine whether an attempt succeeds. SWE-Marathon requires models to complete larger software projects, subsequently running project tests to verify the deliverables.

Both models score higher than Opus 5.5 and Astra on these two long-project benchmarks. On DeepSWE, the two Wonder models differ by just 0.2 percentage points, while on SWE-Marathon, Full surpasses the standard version by 4.4 percentage points.

In SWE-Marathon, Full uses the xhigh reasoning effort level, whereas Astra and Fable use agent environments such as Codex and Claude Code, respectively. Most other results use the max reasoning effort level.

Performance varies across other engineering tasks. Terminal-Bench 4.0 inspects final environment states within isolated terminals; ProgramBench asks models to reconstruct full programs without internet access, requiring each attempt to pass at least 95% of hidden behavioral tests to qualify.

On terminal tasks, Opus and Astra score higher than Full; on program reconstruction, Opus's 70.5% substantially exceeds Full's 59.0%.

The official kitchen demo applies this development capability to a small application: writing code from customizable image inputs to produce a playable 3D kitchen mini-game. The model natively outputs text; the game is executed by the runtime environment via the generated code.

Aggregate Evaluations: Why do the two composite indices rank differently?

Composite indices aggregate performance across multiple tasks. Meta's internal Personal Superintelligence Index (PSI) spans seven personal and knowledge-work benchmarks, covering active focus, business workflows, professional tasks, browsing, and research. The third-party Artificial Analysis Intelligence Index aggregates a different suite of workplace, coding, science, and general reasoning tests.

Full leads on PSI, but falls behind Opus on the AA composite index. The two rankings are not contradictory: differences in task selection and aggregation methods emphasize different priorities.

Can they adapt flexibly to unfamiliar tasks?

Fixed benchmarks cover only limited tasks, while real-world goals, tools, and requirements constantly change. Even when two models achieve similar scores on standard benchmarks, they may diverge when encountering unfamiliar tasks. The discussion around "benchmaxxing" focuses specifically on this gap: optimizing for fixed benchmarks does not necessarily translate into equal real-world adaptability.

Meta compared Muse Spark 1.3 with Wonder Full, scaling inference compute to observe how both models perform on tasks not specifically targeted during post-training.

How to read these two lines?

Dark blue is Wonder Full, light blue is Spark 1.3. The original chart compares both models across different tasks.

View full comparison Horizontal axis: Post-training targeted tasks Vertical axis: Non-targeted post-training tasks View gap at similar horizontal scores

Select a button above to inspect the horizontal axis, vertical axis, and the gap between the two trajectories.

In the chart, the horizontal axis represents performance on tasks heavily targeted during post-training, while the vertical axis represents performance on tasks not specifically optimized during post-training; each point corresponds to a different inference compute budget. Both trajectories trend upward as compute increases; at comparable horizontal-axis performance, Wonder Full reaches higher on the vertical axis—meaning it performs better on this set of non-targeted tasks.

Meta attributes this improvement to a larger, more capable base model, making capabilities acquired during post-training easier to transfer.

Personal agents will encounter novel combinations of schedules, preferences, tools, and ad-hoc changes that cannot be fully anticipated during training; they must apply existing capabilities to these combinations to handle new requirements.

Application Demos: From personal records to concrete deliverables

Official demos show models accessing personal records to deliver routes, orders, seating charts, or follow-up emails via browsers and applications. The preferences, constraints, and changes across these tasks draw upon the memory, execution, and domain understanding capabilities detailed above.

Running: Route planning informed by health habits and schedules

When planning a Saturday long run, the model extracts three constraints from calendar and health records: increasing the long-run distance by 1 km each week, sticking to flat terrain to protect knees, and arriving at the library before 10 AM. Based on these, it maps a 15.1 km route departing at 8 AM, adds markers every 5 km on a map powered by OpenStreetMap data, generates a shareable link, and schedules the run with the link directly into the calendar.

Gifts: Designing and ordering a T-shirt from cat photos

A user wants a birthday T-shirt featuring their cat, Coco. The model browses Coco's Instagram profile, selects actual photos, designs a vintage band-style T-shirt with gold "COCO" lettering according to preferences, proceeds through checkout, and completes order confirmation.

Wedding: Seating guests by requirements with item-by-item verification

The user provides a 47-guest list annotated with accessibility needs, dietary allergies, and seating separations. The model inspects the venue floor plan, converts notes into rules, seats guests table by table in a planning tool according to those rules, and finally cross-checks against original constraints item by item before exporting the seating chart.

Insurance: From bill verification to email follow-up

The user uploads photos of an outpatient bill along with the front and back of an insurance card. Cross-referencing the documents, the model explains billing discrepancies and why insurance was not applied, proposes solutions, and drafts talking points for a call with the clinic's billing department. When the user asks to follow up in writing instead, it adapts by drafting and sending an email, ultimately leading the clinic's billing department to submit the claim to the insurer.

Health: Connecting travel and symptom clues

The model identifies travel and symptom clues across calendar events, messages, and emails, requests permission before connecting to health services such as Withings and HealthEx, and connects altitude data with symptom timing to help the user understand the situation.

Delegation requires strict permission controls

When personal agents process emails, calendars, and private documents, they may encounter external content embedded with malicious instructions. In Meta's evaluations, Full recorded a personal data violation rate of 2.2% and a prompt injection success rate for unauthorized code execution of 10.7%. Actions such as making payments, sending external communications, and modifying or deleting data still require system permission boundaries and explicit user confirmations.

Meta reinforced refusals for chemical, biological, radiological, and nuclear (CBRN) risks as well as cyber misuse, while preserving compliance on legitimate cyber defense tasks. Full refused 98.7% of high-risk biological requests in the BioTIER English evaluation and complied with 93.3% of legitimate cyber defense requests in the defense evaluation. These measure rejecting dangerous uses and assisting legitimate defense, respectively.

According to Meta's Superintelligence Scaling Framework, absent mitigations, it cannot be ruled out that Full reaches high-risk thresholds in chemical and biological domains; after incorporating safety post-training and system-level classifiers, Meta assesses residual risk across all domains as "moderate or lower."