Muse Wonder Official Specifications, Pricing & Benchmarks
Meta Superintelligence Labs flagship Muse Wonder ($4/$20) with 1.05M context and 943K max output.
📋 Muse Wonder Official API & System Card Specifications
Verified against official Meta developer documentation & pricing
1. Identity & Capabilities
API Spec- API Model ID
- muse-wonder
- Release Date
- 2026-10-12
- Weight License
- Proprietary (Undisclosed)
- Native Reasoning
- ✅ Supported
- Function Calling & JSON Schema
- ✅ 原生支持
- Batch API 异步批处理
- ✅ Supported (50% off)
2. Context, Output & Modalities
Limits- Max Context Window
- 1,049,000 tokens (1049K)
- Max Output Tokens
- 943,718 tokens
- Input Modalities
- Text + Image + Video + PDF + Audio
- Output Modalities
- Text
- Direct China Access
- 🌐 Global / OpenRouter
- Deployment Mode
- Cloud API only (Proprietary)
3. Official Pricing & Cache
USD / 1M- Input / 1M Tokens
- $ 4
- Cached Input / 1M
- $ 0.2 -95%
- Output / 1M Tokens
- $ 20
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
📊 Muse Wonder Full Benchmark Scorecard
16 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 4 items
DeepSWE 1.1 evaluates software engineering agents using the mini-swe-agent harness.
SWE-Marathon is an ultra-long-horizon software engineering benchmark covering tasks such as building compilers, optimizing kernels, and developing production-grade services. It measures whether agents can sustain quality across extremely long engineering trajectories.
Artificial Analysis reported accuracy on terminal tasks with executable final-container-state verifiers.
Program Bench evaluates code-generation agents by asking them to recreate a program's behavior from only a compiled binary and documentation.
🤖 Agents & Tools 9 items
Internal recall evaluation over 64K-100K-token conversations containing distractors, conflicting facts, and later updates; LLM judge with weighted question-specific rubrics.
Partial-credit computer-use score on OSWorld 2.1, as reported in the Claude Sonnet 5.5 release. Separate from OSWorld 2.0 and strict completion scores.
JobBench evaluates agents on work-related tasks, as reported in the MiMo-V2.6 release.
Vals Finance Agent v2 evaluates AI agents on realistic financial analyst tasks.
Internal composite of Agents' Last Exam, AutomationBench-AA, FocusBench, JobBench, GDP.pdf, VBT, and DRACO.
Internal personal agent evaluation across housing, travel, IT, and finance; persistence through blockers and adaptation across interactions and changing environments.
AutomationBench-AA is Artificial Analysis's independently run version of AutomationBench, covering 657 real-world SaaS workflow tasks across 40 simulated applications (e.g. Gmail, Slack, Salesforce, HubSpot).
Strict pass rate on OSWorld 2.1: fraction of tasks where every graded checkpoint is satisfied. Separate from the OSWorld 2.1 partial-credit score and from OSWorld 2.0.
Harvey LAB-AA evaluates model performance on complex legal workflows.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
Closest Self-Hostable Open-Weight Alternatives to Muse Wonder
Open-weight models with full HuggingFace weights for private GPU / Mac deployment
Frequently Asked Questions
Is Muse Wonder open-source or proprietary?
Muse Wonder (muse-wonder) is a proprietary commercial model by Meta. Weights and internal layer dimensions are undisclosed; it is accessed via cloud API.
What are the context limit and API pricing of Muse Wonder?
Muse Wonder supports 1049K input context and 943,718 max output tokens, priced at $4/1M input ($0.2/1M cached) and $20/1M output.