MENLO PARK, CALIFORNIA / RankWire.AI / – Meta has launched its most advanced artificial intelligence models yet, named Muse Wonder and Muse Wonder Full. Designed specifically for coding, long-duration AI agents, and intricate personal or professional tasks, these systems aim to assist users in managing insurance claims, interpreting leases, analyzing financial data, and developing functional applications. Muse Wonder was made available on October 12 via the Meta Model API and Muse Code, while Muse Wonder Full was released for subscribers of Muse Code.
This launch marks Meta’s expansion into AI systems capable of handling multi-step processes rather than just responding to single prompts. Muse Wonder can organize extended tasks, assign subtasks to sub-agents, utilize software tools, and recover from errors during execution. The models are also optimized for advanced computing tasks across browsers, desktop programs, and mobile apps. Both models support a context window of approximately one million tokens, enabling them to process large volumes of information within a single session.
To enhance performance on lengthy and personalized tasks, Meta integrated two features called Active Memory and Active Focus. Active Memory retains relevant information across conversations and retrieves details pertinent to new tasks. Active Focus allows the models to continue working through difficult tasks, test alternative approaches, and recover from errors. The models can also ask users for clarification and verify details before executing critical actions, according to Meta.
Models Focus on Coding and Autonomous AI Agents
Using the Meta Model API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens. The company’s developer documentation states that cached input is priced at $0.20 per million tokens. Muse Wonder Full is exclusively available through paid Muse Code subscriptions, not as a separate API option. Developers with higher workloads can opt for Power Usage at $50 a month and Maximum Usage at $200 a month.
Meta reports that Muse Wonder Full outperforms many benchmarks related to agentic and coding capabilities, as shown in results published alongside the launch. Notable tests include DeepSWE and SWE-Marathon, which evaluate software engineering and long-duration coding performance. An independent evaluator, Artificial Analysis, assigned Muse Wonder an Intelligence Index score of 52, and Muse Wonder Full scored 56 at maximum reasoning effort. These scores place the models above Meta’s earlier Muse Spark generation on the evaluator’s scale.
Enhanced Safety Measures Accompany Expanded Model Features
Meta has trained the Muse Wonder series to prevent sharing private data without authorization and to decline requests involving harmful activities. Additionally, the models include cybersecurity functions that assist developers in identifying vulnerabilities within their code and bolstering security defenses. Meta states it assesses its most capable models under the Meta Superintelligence Scaling Framework, with plans to publish a comprehensive safety report covering further safety evaluations following the initial release.
The company has been overhauling its AI efforts for approximately 14 months prior to the Muse Wonder launch. Meta notes that these new models expand upon the foundations laid with Muse Spark, with training conducted at its Prometheus data center. The company intends to integrate Muse Wonder into more features within its Muse ecosystem, beginning with subscribers. The October rollout provides immediate access to Muse Wonder for developers, while reserving the more comprehensive Muse Wonder Full for Muse Code subscribers.