Meta Launches Muse Wonder: Frontier Model for Agents and Coding Available to Developers Now

Meta released its next-generation large model, Muse Wonder, on October 12. The lineup includes two versions: Muse Wonder, which prioritizes speed, and Muse Wonder Full, the complete version. Zuckerberg stated that this model places Meta in the "frontier" tier, putting it on par with the strongest models from OpenAI, Anthropic, and Google.

The model was developed by Meta Superintelligence Labs. According to Meta, the company rebuilt its AI business from the ground up 14 months ago, reconstructing the underlying technology stack. The first result of this effort was the Muse Spark series. Muse Wonder is significantly larger than Muse Spark and was trained at Meta’s gigawatt-scale Prometheus data center.

[Who Can Use It and How]

Developers can access Muse Wonder today via the Meta Model API (Meta’s large model interface platform) or through Meta’s AI coding tool, Muse Code.

The full version, Muse Wonder Full, is currently available only to Muse Code subscribers. The official blog states that it is accessible within Muse Code but does not mention API availability.

For general users, the AI assistant Muse will integrate Muse Wonder "soon," starting with subscribers. However, there is a discrepancy in messaging: Zuckerberg stated on X that "subscribers can use Muse Wonder to power their Muse today." Regarding free users, Meta only said it would gradually bring this level of capability to everyone and plans to develop specialized models based on Muse Wonder for all Muse users in the future, without providing a timeline.

Model card specifications indicate a context window of 1 million tokens, sufficient to ingest a medium-to-large code repository in one go. The model accepts text, images, video, and audio as input, but outputs only text.

[How Does It Compare to Claude and GPT?]

The following results come from comparison tables published by Meta. The competitors are Claude Opus 5.5, Claude Fable 5.1, and GPT-6 Astra, all tested at maximum reasoning intensity. Note that the Personal Superintelligence Index, MemoryBench, and FocusBench are proprietary tests created by Meta. Some competitor scores were run by Meta itself, and the model card explicitly notes that these runs may not have been optimized for the competing models.

Muse Wonder Full leads primarily in three categories of tasks.

First, agentic tasks that act on behalf of users. On Meta’s proprietary Personal Superintelligence Index, Muse Wonder Full scored 59.1, compared to 55.2 for Claude Opus 5.5 and 54.5 for GPT-6 Astra. On AutomationBench, which tests end-to-end business workflows, it scored 45.4 versus the highest competitor score of 41.6.

Second, long-duration coding. On DeepSWE, it scored 79.0 against the highest competitor score of 74.2. On SWE-Marathon, which involves large software projects, it scored 57.5 versus the highest competitor score of 46.3.

Third, legal tasks. On an agentic test by legal AI company Harvey, it scored 13.1 against the highest competitor score of 8.6. This test requires every deliverable item to meet standards, resulting in low pass rates across all models.

However, Muse Wonder Full trails in many areas, mostly losing to Claude Opus 5.5: Terminal-Bench (terminal programming) 57.1 vs. 59.6; ProgramBench (writing a complete program from scratch) 59.0 vs. 70.5; OSWorld (computer operation) 78.4 vs. 81.8; and health Q&A 62.3 vs. 65.6. On GDP.pdf, a professional document task, it lost to GPT-6 Astra, scoring 33.2 vs. 34.2.

Meta’s blog also included Artificial Analysis’s composite intelligence index: Claude Opus 5.5 ranks first with 58 points, while Muse Wonder Full ranks second with 56 points. The version currently available via API for developers is Muse Wonder, which scores 52 points, lower than Claude Fable 5.1, GPT-6 Astra, and Google’s Gemini 4 Argon (all scoring 53 points).

Based on Meta’s data, Muse Wonder excels at letting AI work autonomously for extended periods and remembering user context across conversations. In terms of overall capability, however, it still ranks behind Claude Opus 5.5.

[What Are Active Memory and Active Focus?]

These are names Meta has coined for two key capabilities, which serve as the main selling points of this release.

Active Memory: The ability to remember important information relevant to you across conversations and proactively retrieve applicable past context when performing new tasks. Meta tested this capability using its proprietary MemoryBench. The test involves long conversations ranging from 64,000 to 100,000 tokens, interspersed with distractors and contradictory information that is later updated, checking if the model correctly recalls key details. Muse Wonder scored 85.6, Muse Wonder Full scored 85.5, and Claude Opus 5.5 scored 81.5.

Active Focus: The ability to persistently pursue difficult goals, switching strategies when one path fails, troubleshooting obstacles independently, and adjusting to changing circumstances. It asks questions when necessary and confirms with the user before performing critical operations. Meta tested this using its proprietary FocusBench, featuring tasks related to renting housing, travel, IT, and financial management, requiring the model to navigate various real-world barriers through multi-turn interactions.

In Meta’s demos, Muse Wonder directly operated browsers and multiple apps to generate a Saturday running plan based on the user’s personal preferences and situation.

[What Was Not Disclosed Officially]

↳ Quoting @finkd (Mark Zuckerberg): Introducing Muse Wonder, our new family of frontier models with state-of-the-art performance for personal and professional agentic work.