Meta Releases Muse Wonder Frontier AI Models for Agentic Tasks

Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, describing them as its most capable models yet for personal and professional agentic work. The models are available now through the Meta Model API and Muse Code. Muse Wonder Full is limited to Muse Code subscribers.

Meta said the models were trained at its gigawatt-scale Prometheus data center and build on the pretraining and reinforcement learning approach used for the earlier Muse Spark series. They support a one-million-token context window and accept text, images, video, audio, and PDFs. Pricing for Muse Wonder through the API is $4 per million input tokens and $20 per million output tokens.

The company highlighted two features: Active Memory, which carries relevant details across conversations, and Active Focus, which keeps the model working toward a goal, tries alternative approaches, and recovers from errors. Meta said the models plan multi-step tasks, delegate to sub-agents, and check in before high-stakes actions.

On the Artificial Analysis Intelligence Index, Meta reported Muse Wonder Full at 56 and Muse Wonder at 52, up from 48 for Muse Spark 1.3. The company said the models match or exceed Claude Fable 5.1 and GPT-6 Astra on several benchmarks and place near Claude Opus 5.5. Independent checks showed strong results on coding benchmarks such as DeepSWE and SWE-Marathon. Meta noted that MemoryBench scores should be viewed with caution because the company created that test.

Chief AI Officer Alexandr Wang has described earlier Muse models as competitive with leading closed systems. The new release follows months of work after Meta reorganised its AI efforts. The models will soon power the Muse agent for subscribers and are positioned for tasks ranging from insurance claims and lease review to building working applications.

Meta’s model card lists higher scores in legal, finance, and health domains compared with previous releases. The company has not released full training data details or exact parameter counts. Access remains tiered: the standard Wonder model is available via API and Muse Code, while the Full variant stays inside the coding product for now.

The release arrives as several labs push larger agentic systems that maintain state over long sessions. Meta’s emphasis on memory and error recovery addresses common failure modes in earlier agent demos. Developers can test the models today through the announced channels.

No independent third-party audits of the full benchmark suite have been published yet. Meta said the models represent the next step on its scaling ladder toward personal superintelligence features inside its products.