Two AI models from Meta — Muse Wonder and Muse Wonder Full — have entered the field as the company pushes further into high-end artificial intelligence development. Built within Meta's Superintelligence Lab, the new models target more complex workloads, among them writing and debugging code, as well as retaining details about individual users across interactions.
That product line now has a companion in Muse Spark, a stripped-down version engineered to prioritise speed over depth. According to internal evaluations conducted by Meta, the latest models posted stronger benchmark scores and matched the performance of Fable 5.1 and GPT-6 Astra.
Framed around what the firm terms "personal superintelligence," the launch places particular weight on AI assistant tools capable of retaining and drawing on details from one conversation to the next. Among the capabilities Meta has drawn attention to is what the company calls "active memory" — a function designed to hold onto and connect details a user shares across days, weeks, and months.
According to internal testing carried out by the company, Muse Wonder posted a higher score on MemoryBench — a benchmark tool built to assess continual learning and memory construction — than three rival models, Fable, Opus, and Astra. According to Kavya Kopparapu, a research scientist, the model had received considerable investment from Meta toward retaining details that users find personally relevant across longer timeframes.
Depending on how conditions evolve, AI assistants drawing on that capability could take on a broader range of tasks within longer workflows, including laying out project plans and adjusting recommendations along the way. Those figures come from testing Meta conducted and published internally, leaving open the question of how the models will actually behave when put to work in real-world conditions.
Developers tapping into the Meta Model API will find Muse Wonder priced at $4 for every million tokens fed in and $20 for every million tokens returned. Tokens flowing into the model carry the prompts and data it receives, whereas those flowing out represent the text it produces in response.
Access to Muse Wonder Full through the API has not been made available. Muse Code subscribers will be the only users with access. Meta's Muse agent is set to draw on both models as its underlying infrastructure, the company has stated.
The release marks the first frontier-model deployment to come out of Meta's AI division since Alexandr Wang — who built Scale AI before joining the company — completed a structural reorganisation of that unit in August of the prior year. According to the company, those models drew on the Prometheus data centre in central Ohio — a facility that began operations earlier this year — for their training.
With the rollout, Meta finds itself in more direct rivalry with AI firms including OpenAI and Anthropic. Whether those advertised improvements in memory, coding, and complex reasoning will deliver tangible value for developers and everyday users has not yet been established.
