Quick Facts
- Meta released two frontier models, Muse Wonder and Muse Wonder Full, on Oct. 12, 2026, under its Superintelligence Labs division.
- Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens via commercial API; Muse Wonder Full is locked behind Muse Code subscriptions starting at $50 per month.
- On the Artificial Analysis Intelligence Index, Muse Wonder Full scores 56 and the API-facing Muse Wonder scores 52, compared to Claude Opus 5.5 at 58.
Meta released two new AI models on Monday, Oct. 12, 2026, claiming to have reached the frontier of AI performance after more than a year of rebuilding its research operations and spending tens of billions of dollars. The models, codenamed Watermelon during development, are called Muse Wonder and Muse Wonder Full.
CEO Mark Zuckerberg called the Wonder lineup "a much larger and more advanced class of models" compared to the earlier Muse Spark family. Alexandr Wang, head of Meta Superintelligence Labs, told Axios the models are "true frontier models, competitive with the best models in the world."
Benchmark Results
Independent scores from Artificial Analysis's Intelligence Index place Claude Opus 5.5 at 58, Muse Wonder Full at 56, and the API-accessible Muse Wonder at 52. GPT-6 Astra and Claude Fable 5.1 both sit around 53 on the same index.
The predecessor model, Muse Spark 1.3, scored 48 on that index. The biggest gains appear in agentic evaluations. On Terminal-Bench 4.0, performance jumped from 33% to 57%.
On Meta's own 14-benchmark scorecard, Muse Wonder Full leads in long coding and legal agent tasks. Meta reports Wonder Full at 79.0 on DeepSWE 1.1 versus roughly 74 for Claude Opus 5.5 and GPT-6 Astra. On Harvey's legal agent benchmark, Meta scores Wonder Full at 13.1 against 4.2 for Opus 5.5. Claude Opus 5.5 still outperforms Wonder Full on HealthBench Pro, OSWorld computer use, Terminal-Bench, and ProgramBench, where Opus 5.5 scores 70.5 against Wonder Full's 59.0.
New Capabilities: Active Memory and Active Focus
Both Wonder models include two new features Meta calls Active Memory and Active Focus. Active Memory carries context across conversations and retrieves relevant details during tasks. Active Focus describes the model's ability to sustain effort on complex goals, try alternative approaches, and adapt when it hits obstacles.
Research scientist Kavya Kopparapu said in a video accompanying the launch, "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." On MemoryBench, Muse Wonder scores higher than Claude Opus 5.5, GPT-6 Astra, and Claude Fable 5.1, according to Meta's data.
Pricing and Access
Meta priced Muse Wonder at $4 per million input tokens and $20 per million output tokens, placing it alongside GPT-6 Sol and Claude Opus 5.5. Cached input drops to $0.20 per million tokens. Built-in web search costs $2.50 per 1,000 queries. Rate limits sit at 3,000 requests and 4 million tokens per minute. The model supports a 1 million token context window and accepts text, images, video, and audio.
Muse Wonder is OpenAI SDK compatible. Developers can switch an existing project by changing the base URL and setting the model to muse-wonder-1.0.
Muse Wonder Full is not available via a pay-as-you-go API. Access requires a Muse Code subscription: Power Usage at $50 per month includes Wonder Full with 20 times the standard usage allowance, while Maximum Usage at $200 per month provides 80 times the standard allowance.
This is a notable departure from Meta's history with the Llama family, which made downloadable model weights a standard offering. Neither Wonder model ships with public weights.
Enterprise Considerations
Enterprise buyers evaluating a non-OpenAI, non-Anthropic frontier API endpoint only get access to Muse Wonder, not the top-scoring Wonder Full. Muse Code subscriptions default to Contributor mode, which permits Meta to train on submitted content unless users manually disable the setting. That is a material caveat for organizations handling sensitive client or proprietary data.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.