Meta Launches Muse Wonder, Its First Frontier AI Model, to Rival OpenAI and Anthropic

Meta has officially entered the frontier AI race. On October 12, 2026, the company unveiled Muse Wonder and Muse Wonder Full, the first frontier-class models from Meta Superintelligence Labs and the most capable AI systems Meta has ever shipped. The launch marks the culmination of a 14-month overhaul of Meta’s AI efforts, one that CEO Mark Zuckerberg says has taken the company from rebuilding its foundations to pushing the frontier.

“It’s been just over a year since we started Meta Superintelligence Labs,” Zuckerberg wrote on Threads. “Muse Spark was our first family of models, and now Muse Wonder is a much larger and more advanced class of models that brings us to the frontier.”

Where Muse Spark was built as a lightweight model tuned for low-latency responses, the Wonder models are bigger, score higher on benchmarks, and are designed for complex agentic work such as coding and multi-step task execution. “Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence,” Meta said in its launch blog post.

Active Memory and Active Focus

The headline features are aimed at making AI assistants useful over long stretches of time rather than single conversations. Meta calls the first one Active Memory: the models retain and synthesize what users tell them across days, weeks, and months. To back the claim, Meta highlighted a benchmark called MemoryBench, which tests continual learning and memory construction, and says Muse Wonder now outscores Anthropic’s Fable and Opus models as well as OpenAI’s Astra on it.

“We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond,” said research scientist Kavya Kopparapu in a video released alongside the launch.

The second, Active Focus, is about reliability: the model keeps working toward a hard goal through blockers, recovers from errors, and checks in with the user before consequential actions. Per its OpenRouter listing, Muse Wonder plans, delegates to sub-agents, calls tools in parallel, and can operate browsers as well as desktop and mobile applications.

Benchmarks: Meta’s claims vs independent scores

Meta’s own model card puts Wonder and Wonder Full at benchmark scores comparable to Claude Fable 5.1 and GPT-6 Astra. Its charts show Muse Wonder Full scoring 57.5 on SWE-Marathon, Meta’s long-running software benchmark, against 46.3 for Claude Opus 5.5.

Independent testing paints a competitive but slightly more modest picture. Artificial Analysis scored Muse Wonder Full at 56 on its Intelligence Index at max reasoning effort, two points behind Claude Opus 5.5 at 58, with top-three finishes on six of ten evaluations. The standard Muse Wonder landed at 52. Both are a real jump from Muse Spark 1.3 at 48, which shipped only six weeks earlier.

Pricing and availability

Developers can access Muse Wonder through the Meta Model API at $4 per million input tokens and $20 per million output tokens, roughly in line with GPT-6 Sol or Claude Opus 5.5. The model ships with a 1,048,576-token context window, multimodal input covering text, images, video, audio, and PDF, tool calling, and structured outputs. It is already live on OpenRouter, which reports throughput of 71 tokens per second on the Meta provider.

Muse Wonder Full, the variant posting the headline benchmark numbers, will not be offered through the API. It is available only to subscribers of Muse Code, Meta’s AI coding tool. In the Muse personal AI agent app, Wonder rolls out to paying subscribers first, part of Meta’s push to turn its AI investment into subscription revenue. According to CNBC, Muse subscribers will need tiers priced at $20 or $100 per month to access the new models.

Why it matters

The launch is Meta’s answer to a year of aggressive catch-up spending. After Zuckerberg pledged billions of dollars and poached researchers across the industry, the company formed Meta Superintelligence Labs just over a year ago. Both Wonder models were trained at Meta’s gigawatt-scale Prometheus data center.

Investors liked what they heard. Meta shares rose about 4 percent in afternoon trading on Monday to around $750, ending a four-session losing streak, with analysts staying bullish ahead of the company’s October 28 earnings report.

Meta also says the new models include its strongest safeguards against cyber misuse, and publishes model cards covering red-teaming for cybersecurity, biological risk, and persuasion benchmarks.

Open weights move in parallel

Alongside the frontier launch, Meta opened the weights of Muse Glimmer for developers to download on Hugging Face, with integrations for Ollama, LM Studio, llama.cpp, ExecuTorch, and MLX. Zuckerberg also teased that Muse Spark 1.2 weights are coming soon, signaling Meta plans to keep one foot in open models while it monetizes the frontier ones.

FAQ

What is Meta Muse Wonder?

Muse Wonder is Meta’s first frontier AI model, released October 12, 2026 alongside the more powerful Muse Wonder Full. It is built for agentic tasks like coding, planning, and multi-step work, with long-term memory features Meta calls Active Memory and Active Focus.

How does Muse Wonder compare to GPT-6 and Claude?

Meta says Wonder matches Claude Fable 5.1 and GPT-6 Astra on its benchmarks. Independent testing by Artificial Analysis ranks Muse Wonder Full at 56 on the Intelligence Index, two points behind Claude Opus 5.5 at 58.

How much does Muse Wonder cost?

Via the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is not sold through the API and is only available to Muse Code subscribers.

How do I try Muse Wonder?

Developers can use it now through the Meta Model API or OpenRouter. In the Muse app, it rolls out to paying subscribers first.

What is Meta Superintelligence Labs?

The AI research division Zuckerberg formed just over a year ago after an expensive shake-up of Meta’s AI efforts, tasked with building personal superintelligence.

Related: Want free tokens to try it? See our Muse AI referral code guide for how to get 1 billion free Muse tokens.