Meta Has Unveiled the Muse Wonder and Muse Wonder Full AI Models for Autonomously Handling Complex Tasks

• Meta has released two AI models, Muse Wonder and Muse Wonder Full, with support for a context window of up to 1 million tokens.
• The developers reported strong performance in programming, financial analysis, document work, and completing multi-step tasks.
• In the independent Artificial Analysis ranking, the models scored 52 and 56 points, respectively, outperforming the previous Muse Spark 1.3.

Meta unveiled the Muse Wonder family of artificial intelligence models, which includes the eponymous model and its higher-performance version, Muse Wonder Full. According to the developers, these are the company’s most powerful AI systems to date, designed to handle complex user and professional tasks.

The release took place on October 12, 2026. The new models mark the next stage in the evolution of the Muse lineup following the launch of Muse Spark. They were trained using the Prometheus data center’s compute infrastructure, with a primary focus on the AI’s ability to independently carry out long-running tasks and adapt to changing conditions.

Meta highlighted two key technologies. Active Memory enables the models to retain important information from previous conversations and retrieve the necessary context when handling new requests. 

Active Focus is responsible for sustained work on complex tasks, searching for alternative solutions, and correcting errors without the need to constantly receive additional instructions.

In particular, the new systems can analyze lease agreements, work with financial and medical documents, build applications, and perform actions in a browser. 

As examples, the developers demonstrated creating a 3D game, organizing a personalized workout plan, and preparing a wedding guest seating chart.

Muse Wonder is already available to developers via the Meta Model API and the Muse Code environment. The more powerful Muse Wonder Full is only available to Muse Code subscribers at launch. 

Going forward, the company plans to integrate the new capabilities into its own Muse AI assistant, initially providing access to paid users.

Model Specifications and Expert Assessments

According to the technical documentation published by Meta, both models support a context window of up to 1 million tokens. They can process text, images, audio, and video, but generate responses in text format.

In its own tests, the company compared Muse Wonder and Muse Wonder Full with OpenAI’s GPT-6 Astra, as well as Anthropic’s Claude Fable 5.1 and Claude Opus 5.5.

Results from some of the benchmarks:

• in the Harvey LAB-AA test, which evaluates performance on legal tasks, Muse Wonder Full scored 13.1 points versus 8.6 for GPT-6 Astra and 4.2 for Claude Opus 5.5
• in Vals Finance Agent V2, designed to assess financial analysis, the flagship model scored 59.1 points versus 58.6 for Claude Opus 5.5
• in MemoryBench, which tests the ability to retain and retrieve information from long conversations, the standard Muse Wonder scored 85.6 points, while the Full version scored 85.5
• in DeepSWE 1.1, focused on programming, Muse Wonder Full posted 79% versus 74.2% for Claude Opus 5.5
• in SWE-Marathon, which evaluates execution of long-running software projects, the flagship model reached 57.5% versus 46.3% for Claude Opus 5.5.

That said, Meta’s products did not lead in every category. For example, in Terminal-Bench 4.0, Claude Opus 5.5 scored 59.6 points versus 57.1 for Muse Wonder Full, and in HealthBench Pro, designed to evaluate answers to medical questions, Claude scored 65.6 points versus 62.3 for Meta’s solution.

The developers themselves clarified that some of the results are based on internal testing, while some competitor metrics were taken from their published reports. As a result, the testing methodologies are not fully identical in every case.

An independent assessment of the new releases was published by the research platform Artificial Analysis. In its Intelligence Index v4.3 ranking, Muse Wonder scored 52 points, and Muse Wonder Full scored 56. For comparison, the previous Muse Spark 1.3 scored 48 points, while Claude Opus 5.5 scored 58.

Analysts noted that Meta’s flagship model placed in the top three in six out of the ten tests that make up the index. The most notable improvements over its predecessor were recorded in workflow automation, professional tasks, and scientific research.

In particular, Muse Wonder Full’s Terminal-Bench 4.0 score rose from 33% for the previous model to 57%. In the GDPval-AA v2.1 test, the result increased from 1,681 to 1,857 Elo points.

At the same time, Artificial Analysis specialists pointed to the relatively high cost of using Muse Wonder. The company set pricing at $4 per 1 million input tokens and $20 for the same volume of output tokens. Processing cached input costs $0.20 per 1 million tokens.

According to analysts’ calculations, the average cost of completing a single task within the Intelligence Index came to $4.52. That is about 2.8 times higher than Muse Spark 1.3, although still lower than the cost of running Claude Opus 5.5 under a comparable test configuration.

Experts also drew attention to limitations in factual knowledge. In the AA-Omniscience test, Muse Wonder Full scored 30 points versus 46 for Claude Opus 5.5 and 43 for GPT-6 Astra. At the same time, the hallucination rate for Meta’s flagship model was 23%, one of the lowest figures among the advanced systems tested.

Reviewers at WIRED also noted Meta’s significant progress after the less successful releases in the Llama family. In the outlet’s assessment, the company managed to close the gap with leading AI developers in coding capabilities and professional analytics. However, journalists emphasized that the test results still need to be validated through large-scale real-world use.

Analysts at investment firm Baird, led by Colin Sebastian, and at Bank of America, led by Justin Post, view the release of Muse Wonder as a potential long-term growth driver for Meta’s business. 

In their view, developing in-house frontier AI models opens up additional monetization opportunities, both through direct access sales and by integrating the technology into the company’s existing services.

Meta also reported strengthening the safety mechanisms of the new systems. The developers focused on protecting personal information, resilience to malicious instructions, and preventing potentially dangerous uses of the models.

In its technical documentation, the company acknowledged that without additional constraints, it cannot rule out Muse Wonder Full reaching a high level of risk in the area of chemical and biological threats. After implementing safeguards, Meta assessed the residual risks across all categories under review as moderate or lower.

The company plans to publish a full report with additional security audit findings later.

Сообщение Meta Has Unveiled the Muse Wonder and Muse Wonder Full AI Models for Autonomously Handling Complex Tasks появились сначала на INCRYPTED.