Muse Wonder: Meta Omni-Modal Architecture, Benchmarks & API Guide

Executive Summary

An authoritative systems engineering breakdown of Meta's Muse Wonder commercial API model and Muse Wonder Full subscription flagship, covering omni-modal early-fusion ingestion, cross-session semantic memory graphs, SWE-Marathon and DeepSWE resilience, Artificial Analysis and LMSYS Chatbot Arena rankings, and enterprise token economics.

Architectural Overview: What Is Meta's Muse Wonder?

Muse Wonder reached general availability on October 12, 2026, marking Meta Superintelligence Labs' entry into persistent, stateful commercial frontier intelligence. Rather than shipping a single monolithic endpoint, Meta structured the release across two complementary flagship variants: the commercial Muse Wonder API accessible via the Meta Model API and OpenRouter, and Muse Wonder Full, an expanded compute tier reserved exclusively for subscribers of Meta's Muse Code Power ($50/month) and Maximum ($200/month) agentic engineering plans. Both tiers share a ground-up omni-modal foundation capable of ingesting interleaved text, high-resolution images, continuous video streams, and multi-speaker audio across a native 1,000,000-token context window while generating up to 128,000 completion tokens in a single pass. What separates Muse Wonder from stateless transformer APIs is the native integration of two stateful cognitive subsystems: Active Memory, which constructs a persistent cross-session semantic memory graph of codebase conventions and institutional facts, and Active Focus, a closed-loop execution controller engineered for autonomous error recovery across multi-hour engineering, computer-use, and regulatory auditing workflows.

Key Takeaways

Dual-Tier Flagship Launch on October 12, 2026

Meta Superintelligence Labs launched Muse Wonder for elastic pay-as-you-go commercial API workloads alongside Muse Wonder Full, an unthrottled flagship exclusive to Muse Code Power ($50/mo) and Maximum ($200/mo) tiers.

Native 1,000,000-Token Omni-Modal Window with 128K Output

Muse Wonder processes text, image, video, and audio inputs directly inside a unified 1,000,000-token context space and emits up to 128,000 structured text tokens without external transcription or frame-captioning glue.

Active Memory Cross-Session Semantic Graph

Through Active Memory, the platform maintains an evolving entity-relationship graph across separate sessions, preserving architectural invariants, user preferences, and verified debugging histories without re-stuffing prompts.

Active Focus Autonomous Error Recovery Engine

Active Focus keeps long-horizon agents locked onto primary objectives during SWE-Marathon, DeepSWE, desktop computer-use, and multi-document audits, automatically backtracking when tool executions or tests fail.

Top-Three Global Frontier Intelligence Rankings

Muse Wonder Full achieves a 56 on the Artificial Analysis Intelligence Index—tying Claude Sonnet 5.5 for #2 globally just behind Claude Opus 5.5 (58)—and ranks #3 on LMSYS Chatbot Arena at ~1509 Elo, while API Muse Wonder scores 52.

95% Cached Input Savings for Enterprise Agent Loops

Priced at $4.00/M input and $20.00/M output tokens with $0.20/M cached input reads, the API tier completes the Artificial Analysis benchmark suite at a $3.27 task cost, matching Gemini 4 Argon and GPT-6.1 Sol on intelligence.

Architectural & Engineering Deep Dive

Unified Omni-Modal Early-Fusion Perception Across 1,000,000 Tokens

Legacy multimodal stacks bolt separate vision towers and speech-to-text frontends onto text backbones, losing temporal alignment between video frames and audio waveforms. Meta Superintelligence Labs built Muse Wonder around a unified early-fusion tokenizer and interleaved spatiotemporal attention stack. Visual frames, spectrograms, UI coordinates, and code tokens project into a shared representation space from the earliest layers, letting cross-modal heads bind a spoken timestamp or acoustic anomaly directly to a video frame and log line hundreds of thousands of tokens away. Dynamic temporal pooling compresses static video regions while preserving full 1,000,000-token recall and a 128,000-token output ceiling.

Active Memory: Persistent Cross-Session Semantic Knowledge Graphs

Stateless APIs suffer from amnesia whenever a context window closes, forcing teams to maintain brittle vector-retrieval pipelines that surface stale chunks. Muse Wonder solves this natively with Active Memory—a transactional semantic graph managed alongside inference. As an agent explores a codebase or document corpus, Active Memory distills verified facts, architectural invariants, API contracts, and user constraints into typed nodes with provenance pointers. When schemas change, Active Memory invalidates superseded edges rather than appending conflicting text, hydrating only relevant subgraphs into future sessions.

Active Focus: Closed-Loop Goal Persistence and Self-Correcting Execution

Long-horizon suites like SWE-Marathon, DeepSWE, and desktop computer-use expose horizon drift, where repeated tool errors push models toward stubbed workarounds. Muse Wonder counters drift via Active Focus, an internal hierarchical objective ledger with counterfactual backtracking. Before running shell commands or edits, Active Focus defines verifiable postconditions; when tests or UI assertions fail, it quarantines the failed branch, summarizes the falsified hypothesis, and resumes from the last verified checkpoint. This discipline drives the Full variant to 56 on the Artificial Analysis Index and #3 (~1509 Elo) on LMSYS Chatbot Arena.

Comprehensive Model Specifications

Head-to-Head Performance & Benchmark Matrix

Real-World Implementation & Hands-on Verification

Multi-Hour SWE-Marathon & DeepSWE Monorepo Migration with Active Focus

Scenario Evaluation: A platform squad runs Muse Wonder Full inside Muse Code Maximum ($200/mo) to migrate a 620,000-token Rust and TypeScript settlement monorepo from legacy gRPC bindings to async tonic v0.14.

Standardized Benchmark Prompt:

Load all workspace crates and protobuf schemas, preserve v1 wire compatibility, run cargo test --workspace after each crate patch, and recover cleanly from borrow-checker or test regressions.

Empirical Output Summary: Across 94 steps, Active Focus caught two async lifetime deadlocks during integration tests, rolled back only the offending trait impls, and stored verified crate invariants in Active Memory before passing all 418 tests.

Evaluation Verdict: Unlike stateless agents that loop on compiler errors, Muse Wonder preserved goal invariants across 3.5 hours of autonomous refactoring.

Omni-Modal Outage Forensics Over Video, Audio, and Telemetry

Scenario Evaluation: SREs feed the Muse Wonder API a 28-minute Grafana screen recording, a 40-minute incident bridge audio file, and 380,000 tokens of Envoy proxy logs.

Standardized Benchmark Prompt:

Correlate spoken deployment timestamps from the incident audio with visual latency spikes on the Grafana video and isolate the Envoy circuit-breaker commit causing 503 errors.

Empirical Output Summary: Muse Wonder aligned a verbal canary rollout at 14:12:08 UTC with a visual p99 latency spike at 14:12:19 UTC in the video, matched the Envoy diff, and emitted a YAML fix in one 1M-token pass.

Evaluation Verdict: Native four-modality fusion saved forty minutes of manual transcription and screenshot stitching during live triage.

Million-Token Clinical, Financial, and Legal Due-Diligence Audit

Scenario Evaluation: A compliance team uses Muse Wonder with cached context ($0.20/M reads) to audit 890,000 tokens of Phase III clinical protocols, FDA letters, and supplier indemnification contracts across twelve sessions.

Standardized Benchmark Prompt:

Cross-examine adverse-event reporting windows in the clinical protocols against liability carve-outs in Schedule 4B, persisting verified entity mappings across sessions via Active Memory.

Empirical Output Summary: The model flagged three reporting-window discrepancies, cited exact clause coordinates, and reused its Active Memory entity graph on follow-up sessions at a 94% cache-hit ratio.

Evaluation Verdict: Pairing $0.20/M cached reads with Active Memory cut multi-turn audit spend by over 80% versus uncached endpoints.

Evidence Ledger: Fact-Check & Verification Audit

To ensure search engine E-E-A-T integrity, claims are classified across confirmed, reported, unverified, and unknown tiers:

Production Caveats & Known Constraints

Muse Wonder Full Exclusive to Paid Muse Code Subscriptions

Organizations building custom third-party SaaS backends cannot currently call Muse Wonder Full over raw pay-per-token REST endpoints, as the 56-AA-Index variant remains exclusive to Muse Code Power ($50/mo) and Maximum ($200/mo) seats.

Closed Proprietary Weights Require Cloud Data Governance Review

Unlike its open-weight sibling Muse Spark Open, Muse Wonder is a proprietary managed cloud service; strictly air-gapped defense, sovereign, or zero-egress environments cannot self-host its weights on private clusters.

Text-Only Generation Output Modality

Although Muse Wonder natively perceives interleaved text, images, video, and audio on the input side, its generation head emits text and structured tool calls only, requiring downstream TTS or diffusion models for media synthesis.

Recommended Developer Action Plan

Provision Meta Model API or OpenRouter Credentials with Prompt Caching

Connect your production agent gateway to the Muse Wonder endpoint on Meta Model API or OpenRouter and enable deterministic prefix and semantic caching to lock in the $0.20/M cached input rate.

Enable Scoped Active Memory Namespaces Per Repository or Tenant

Configure isolated Active Memory tenant namespaces in your SDK initialization so persistent cross-session semantic graphs remain strictly partitioned by customer, repository, or compliance boundary.

Evaluate Muse Code Power ($50/mo) or Maximum ($200/mo) for DeepSWE Workflows

Equip senior staff engineers with Muse Code subscriptions to benchmark Muse Wonder Full (56 AA Index) on multi-hour autonomous repository migrations and desktop computer-use automation.

Frequently Asked Questions

What is Muse Wonder and when did Meta release it?

Muse Wonder is a flagship omni-modal, stateful agentic model family developed by Meta Superintelligence Labs and released into general availability on October 12, 2026 across the Meta Model API, OpenRouter, and Muse Code.

How does Muse Wonder differ from Muse Wonder Full?

Muse Wonder is the commercial pay-per-token API variant scoring 52 on the Artificial Analysis Intelligence Index, whereas Muse Wonder Full is a larger flagship variant scoring 56 (tied #2 globally) offered exclusively inside Muse Code Power ($50/mo) and Maximum ($200/mo) subscriptions.

What input and output modalities does Muse Wonder support?

Muse Wonder natively accepts four input modalities—text, image, video, and audio—within a unified 1,000,000-token context window and generates structured text outputs up to 128,000 completion tokens.

How do Active Memory and Active Focus improve agentic reliability?

Active Memory builds a persistent cross-session semantic graph so agents retain codebase and business context across separate runs, while Active Focus enforces closed-loop goal tracking and autonomous error recovery during SWE-Marathon, DeepSWE, and computer-use tasks.

How does Muse Wonder rank on Artificial Analysis and LMSYS Chatbot Arena?

Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index—tying Claude Sonnet 5.5 for #2 globally right behind Claude Opus 5.5 (58)—and ranks #3 on LMSYS Chatbot Arena (~1509 Elo). The API Muse Wonder model scores 52, matching Gemini 4 Argon and GPT-6.1 Sol.

What is the API pricing for Muse Wonder?

On the Meta Model API and OpenRouter, Muse Wonder costs $4.00 per million uncached input tokens, $20.00 per million output tokens, and $0.20 per million cached input tokens, yielding a $3.27 Artificial Analysis benchmark task cost.

Can enterprises self-host Muse Wonder on private GPU clusters?

No, Muse Wonder is a proprietary managed model served via cloud endpoints and Muse Code. Organizations requiring self-hosted open weights on private vLLM or SGLang clusters should deploy Meta's sibling release, Muse Spark Open (816B-A42B).

Which enterprise workloads benefit most from Muse Wonder?

Muse Wonder excels at multi-hour repository refactoring (DeepSWE and SWE-Marathon), audiovisual incident root-cause analysis, GUI computer-use automation, and million-token legal, financial, and clinical regulatory document audits.

Verified Sources & References

1. [Meta Superintelligence Labs] Introducing Muse Wonder and Muse Wonder Full: Persistent Omni-Modal Intelligence (ID: src-meta-muse-wonder-announce)
2. [Meta Developer Platform] Meta Model API Documentation: Muse Wonder Endpoints, Context Caching & Pricing (ID: src-meta-model-api-docs)
3. [Meta Muse Code Documentation] Muse Code Power ($50/mo) and Maximum ($200/mo) Subscriptions & Muse Wonder Full Guide (ID: src-muse-code-docs)
4. [Artificial Analysis] Muse Wonder & Muse Wonder Full Intelligence Index, Task Cost, and Latency Evaluation (ID: src-aa-muse-wonder)
5. [LMSYS Org / Chatbot Arena] LMSYS Chatbot Arena Leaderboard (October 2026): Muse Wonder Full ~1509 Elo Ranking (ID: src-lmsys-arena-oct2026)
6. [OpenRouter] Meta: Muse Wonder Commercial API Routing, Throughput & Cache Rate Card (ID: src-openrouter-muse-wonder)