Editorial illustration for Meta Muse Wonder vs Claude: Rival Runs Best-Effort, Not Common Harness
Meta's Muse Wonder Rivals Claude for Long-Run Tasks
Meta Muse Wonder vs Claude: Rival Runs Best-Effort, Not Common Harness
Meta put out two new models on Monday, Muse Wonder and Muse Wonder Full, and the pitch is straightforward: these are the company's most capable systems yet, built specifically for agents that run for hours on coding, legal work, finance tasks, and computer use rather than quick chat replies. Muse Wonder is already live on the Meta Model API and through Muse Code, Meta's own command-line tool for macOS, Linux, and Windows. Muse Wonder Full is held back for paid Muse Code subscribers only, and both run exclusively on Meta's cloud, no local or third-party hosting option.
The specs point to serious ambition: a 1-million-token context window, input across text, image, video, and audio, output in text. Meta trained Wonder at its Prometheus data center, crediting a bigger base model for improved handling of unfamiliar tasks. Two features get top billing in Meta's framing, Active Memory and Active Focus, both aimed at making the model behave less like a tool and more like a persistent worker.
Benchmark numbers tell a mixed story. Muse Wonder Full tops 6 of 14 rows against Claude Opus 5.5, and trails on several others, including a rough showing on ProgramBench.
Meta has released Muse Wonder and Muse Wonder Full, its most capable models so far. Muse Wonder is a frontier model from Meta built for personal and professional AI agents. Both models target long-running agent work: coding, legal and finance tasks, and computer use.
Why this matters
The 1M-token context and multimodal input make Muse Wonder a serious entry for agent work, but the benchmark asterisk matters more than the specs. Meta's own model card admits rival runs were best-effort, and the SWE-Marathon numbers pit Muse Wonder against Claude Opus 5.5 and GPT-6 Astra using Codex and Claude Code rather than a shared harness. That's not a scandal, it's just not proof of superiority, and we'd treat any chart built on it the same way.
For developers deciding where to route coding or finance agents, the practical test is still your own eval suite on your own tasks, run on Meta's cloud since that's the only place Muse Wonder Full lives right now. Founders should note the paywall: Muse Wonder Full sits behind paid Muse Code plans, so the "frontier" claim comes with a subscription tag attached.
Watch whether Meta or an outside group publishes a same-harness comparison. Until then, treat these numbers as a vendor's best pitch, not a settled ranking.
Related Reading
Browse all Anthropic news →
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
Related Topic Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps
Related Topic Anthropic: Hackers Stole Claude Tokens Via Compromised Session Key
LLMs & Generative AI
From Archives Square's New Kiosk Uses Conversational AI to Help Customers Place Orders
From Archives Nvidia’s Huang pledges USD 2B for Europe, USD 250M for UK AI safety
From Archives Anthropic Self-Replicating AI Warning: Gemma 4 Split Across 15 Devices
Research & Benchmarks
From Archives OpenAI, Anthropic Tell British Lawmakers: Regulate Us, Focus on Standards
Policy & Regulation
Common Questions Answered
What are the key differences between Muse Wonder and Muse Wonder Full?
Muse Wonder is available to all users through the Meta Model API and Muse Code command-line tool for macOS, Linux, and Windows. Muse Wonder Full is restricted exclusively to paid Muse Code subscribers, offering enhanced capabilities for those willing to pay for premium access.
What specific tasks are Muse Wonder models designed to handle?
Muse Wonder and Muse Wonder Full are built specifically for long-running agent work including coding, legal work, finance tasks, and computer use rather than quick chat replies. These models are optimized for personal and professional AI agents that can run for hours on complex tasks.
Why does the article caution about Muse Wonder's benchmark comparisons against Claude and GPT-6?
Meta's own model card admits that rival runs were conducted on a best-effort basis rather than using a common harness for testing. The SWE-Marathon benchmarks pit Muse Wonder against Claude Opus 5.5 and GPT-6 Astra using different tools (Codex and Claude Code), which means the results don't provide definitive proof of superiority across models.
What technical capabilities does Muse Wonder offer for extended workloads?
Muse Wonder features a 1M-token context window and multimodal input capabilities, making it well-suited for handling extended agent work. These specifications enable the model to process and maintain context over long-running tasks that require sustained reasoning and interaction.
Further Reading
- Meta releases its first frontier model, Muse Wonder - TechCrunch
- Meta Unveils New Series of Advanced A.I. Models - The New York Times
- Meta has some Wonder Full new AI models - The Verge
- Muse Wonder: a material jump in intelligence over previous Meta models - Artificial Analysis
- Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents - Meta
