- Meta’s Muse Wonder adds memory and focus for complex tasks.
- Muse Wonder Full outperforms Muse Spark 1.3 in AI agent benchmarks.
Meta has introduced two new AI models, Muse Wonder and Muse Wonder Full, designed to handle complex tasks involving coding, research, and professional work as it competes with OpenAI and Anthropic.
The models introduce two capabilities called Active Memory and Active Focus. According to Meta, these allow AI agents to retain relevant information across conversations and continue working on tasks that require multiple steps.
Muse Wonder is available through Meta’s Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. Meta also plans to introduce Muse Wonder into its Muse AI agent, beginning with subscribers.
How Muse Wonder handles complex tasks
Active Memory allows the models to retain relevant information across conversations and retrieve it when needed. Meta said this reduces the need for users to repeatedly provide information about their preferences, circumstances, and previous interactions.
Active Focus addresses the models’ ability to continue working toward a goal. Meta said the models can try different approaches, work through obstacles, adjust to new information, and request clarification when needed. They are also designed to check with users before taking important actions.
In one demonstration, Meta showed Muse Wonder working across applications to prepare a running plan based on a user’s preferences and circumstances. Other demonstrations involved creating a wedding seating chart with multiple requirements and designing customised merchandise using photographs from an Instagram profile.
Meta said the models can assist with tasks such as understanding lease agreements, interpreting financial information, and handling insurance claims. Its evaluations cover professional applications in law, finance, and healthcare, including tasks requiring the interpretation of documents and detailed instructions.
Software development is another area of focus. Muse Wonder can work across large codebases, make changes to multiple files, and carry out testing and revisions. It also supports the use of sub-agents to divide work into smaller tasks that can run in parallel.
The release builds on Meta’s earlier Muse Spark family. Muse Spark was designed as a lightweight model with low latency, while Muse Wonder and Muse Wonder Full are larger models intended to handle more demanding tasks.
Both Muse Wonder models retain a context window of approximately one million tokens, which was already supported by Muse Spark 1.3. Meta attributes the improvements to larger base models and changes in pretraining and reinforcement learning.
Meta also reported stronger performance on unfamiliar tasks that were not specifically targeted during post-training. The company said the models were trained using its Prometheus data centre infrastructure.
How Muse Wonder compares with rival models
Meta published benchmark comparisons with competing AI models, including GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1. Its evaluations cover professional knowledge, coding, memory, and sustained task execution.
Independent testing by Artificial Analysis also found improvements over Meta’s previous models. In its Intelligence Index version 4.3, Muse Wonder achieved a score of 52, while Muse Wonder Full scored 56 at maximum reasoning effort. Muse Spark 1.3 scored 48 under the same testing conditions.
The largest improvements were recorded in evaluations involving AI agents. On Terminal-Bench 4.0, which tests agents completing tasks in terminal environments, Muse Wonder Full scored 57%, compared with 33% for Muse Spark 1.3. The standard Muse Wonder model scored 52%.
Artificial Analysis also found improvements in coding performance. Muse Wonder Full scored 68 on the Artificial Analysis Coding Agents Index at maximum reasoning effort, compared with 54 for Muse Spark 1.3. The result matched Claude Code using Claude Sonnet 5.5 under the same evaluation setting.
Muse Wonder Full ranked among the top three models in six of the ten evaluations included in the Artificial Analysis Intelligence Index. It performed well in agent-based knowledge work and scientific reasoning, although competing models recorded higher scores in other categories.
Meta also tested the models’ ability to maintain focus during complex tasks. Its FocusBench evaluation assesses sustained effort when working through challenging problems, while SWE-Marathon measures performance on long-running software engineering projects.
Pricing, availability, and data protection
Meta has priced Muse Wonder at US$4 per million input tokens and US$20 per million output tokens through its Model API. Cached input tokens cost US$0.20 per million, according to the company’s developer documentation.
The pricing is higher than Muse Spark 1.3, which costs US$1.25 per million input tokens and US$4.25 per million output tokens. Artificial Analysis found that Muse Wonder cost an average of US$4.52 per Intelligence Index evaluation task, compared with US$1.60 for Muse Spark 1.3.
The independent evaluator attributed the higher cost to Muse Wonder’s API pricing and input-token usage. Its testing recorded an average cost of US$3.26 for GPT-6 Astra and US$5.98 for Claude Opus 5.5 under their respective evaluation settings.
Muse Wonder Full is only available through Muse Code subscriptions and is not offered through the pay-as-you-go API. It is a separate model rather than an additional reasoning setting for Muse Wonder, and Meta has not published comparable per-token API pricing for the Full version.
Meta said the models were trained to handle private information from messages, calendars, and documents while resisting attempts to obtain sensitive information without authorisation. The company also said the models are designed to seek user confirmation before taking important actions.
Meta’s developer documentation distinguishes between Standard mode and Contributor mode. Under Standard mode, the company says submitted prompts, files, code, and model responses are not used for model training. Contributor mode permits such content to be used for training in exchange for higher usage limits, with an option to disable the setting.
Meta said it trains, evaluates, and releases its models in accordance with its Superintelligence Scaling Framework, which covers risks including cybersecurity misuse. The company plans to publish additional safety evaluation results.
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.

