Meta unveiled Muse Wonder and the higher-performance Muse Wonder Full on October 12, 2026. The company said the models are its most powerful AI systems to date, designed to perform complex professional work, independently complete long-running tasks and adapt to changing conditions.

The models follow Muse Spark and were trained using the Prometheus data center’s compute infrastructure. Meta said Active Memory allows them to retain important information from previous conversations and retrieve relevant context, while Active Focus supports sustained work, alternative approaches and error correction without constant additional instructions.

According to Meta, the systems can analyze lease agreements, process financial and medical documents, build applications and perform browser actions. Demonstrations included creating a 3D game, organizing a personalized workout plan and preparing a wedding guest seating chart.

Muse Wonder is available through the Meta Model API and the Muse Code environment. Muse Wonder Full is initially limited to Muse Code subscribers. Meta plans to add the capabilities to its Muse AI assistant, starting with paid users.

Specifications and benchmark results

Meta’s technical documentation says both models support context windows of up to 1 million tokens. They accept text, images, audio and video as inputs but generate text responses.

Meta compared the models with OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 and Claude Opus 5.5. In the Harvey LAB-AA legal benchmark, Muse Wonder Full scored 13.1, against 8.6 for GPT-6 Astra and 4.2 for Claude Opus 5.5. It scored 59.1 in Vals Finance Agent V2, compared with 58.6 for Claude Opus 5.5.

In MemoryBench, Muse Wonder scored 85.6 and Muse Wonder Full scored 85.5. Muse Wonder Full recorded 79% in the DeepSWE 1.1 programming test, against 74.2% for Claude Opus 5.5, and 57.5% in the long-running software-project benchmark SWE-Marathon, compared with 46.3% for Claude.

Meta did not lead every test. Claude Opus 5.5 scored 59.6 in Terminal-Bench 4.0, compared with 57.1 for Muse Wonder Full, and 65.6 in HealthBench Pro, against 62.3 for Meta’s model. Meta said some results came from internal testing, while some competitor figures came from published reports, meaning the methodologies were not fully identical.

Independent assessment and costs

Research platform Artificial Analysis published an independent assessment that gave Muse Wonder 52 points and Muse Wonder Full 56 in its Intelligence Index v4.3. Muse Spark 1.3 scored 48, while Claude Opus 5.5 scored 58.

Artificial Analysis said Muse Wonder Full placed in the top three in six of the index’s 10 tests, with its largest gains over its predecessor in workflow automation, professional tasks and scientific research. Its Terminal-Bench 4.0 result rose from 33% to 57%, while its GDPval-AA v2.1 score increased from 1,681 to 1,857 Elo points.

Meta priced Muse Wonder at $4 per 1 million input tokens and $20 per 1 million output tokens, with cached input costing $0.20 per 1 million tokens. Artificial Analysis calculated an average cost of $4.52 per Intelligence Index task, about 2.8 times the cost of Muse Spark 1.3 but below Claude Opus 5.5 under a comparable configuration.

Muse Wonder Full scored 30 in the AA-Omniscience factual-knowledge test, compared with 46 for Claude Opus 5.5 and 43 for GPT-6 Astra. Its hallucination rate was 23%, which Artificial Analysis described as one of the lowest among the advanced systems tested.

WIRED said Meta had made significant progress after less successful Llama releases and had narrowed the gap with leading developers in coding and professional analytics, while cautioning that the results require validation through large-scale real-world use.

Analysts at Baird led by Colin Sebastian and Bank of America analysts led by Justin Post said Muse Wonder could become a long-term growth driver, with potential monetization through direct access and integration into Meta’s existing services.

Safety measures

Meta said it strengthened protections for personal information, resistance to malicious instructions and controls against potentially dangerous uses. The company acknowledged that, without additional constraints, it could not rule out Muse Wonder Full reaching a high-risk level for chemical and biological threats. After applying safeguards, Meta assessed residual risks in all reviewed categories as moderate or lower and said it plans to publish a fuller security-audit report.

Source: Incrypted