Meta releases Muse Wonder and Muse Wonder Full

Meta launched Muse Wonder and Muse Wonder Full on Oct. 12, 2026, with split access and disputed parity claims.

On Oct. 12, 2026, Meta released Muse Wonder and Muse Wonder Full, successors to Muse Spark that it calls its most capable models for agent work in law, finance, health, and coding, adding Active Memory, Active Focus, a one-million-token multimodal context, and limits on sharing private data or assisting harmful requests while remaining usable for defensive code review. TechCrunch attributes the pair to Meta’s Superintelligence Lab, calls them Meta’s first frontier models, and says they were trained at the Prometheus data center in Ohio, while Latent Space describes a second-generation family about 14 months after the lab’s founding. Meta cites internal tests against GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1; TechCrunch reports parity with Fable 5.1 and GPT-6 Astra and higher MemoryBench results than Fable, Opus, or Astra, whereas The Decoder and Latent Space give Artificial Analysis Intelligence Index v4.3 scores of 52 and 56 against 48 for Muse Spark 1.3 and 58 for Opus 5.5, and Simon Willison says Artificial Analysis equates Wonder (Max) with GPT-6.1 Sol. Wonder is priced at $4 per million input tokens and $20 per million output tokens on the Meta Model API and in Muse Code, but TechCrunch, The Decoder, and Latent Space say Wonder Full is Muse Code-only and not on the API, while The Verge says both are available to Muse Code subscribers; sources also disagree on whether and when both will power the Muse agent. An Oct. 14 Verge column adds that Google’s Gemini 4 Argon, with introductory pricing up to 80% below rivals, and Muse Wonder Full are claimed to match leading OpenAI and Anthropic models after Muse Spark fell short; a MarkTechPost page supplies no verifiable model details.

• On Oct. 12, 2026, Meta released Muse Wonder and Muse Wonder Full, successors to Muse Spark, with Active Memory, Active Focus, and a 1 million-token context for text, image, video, and audio.
• Wonder is on the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens; Artificial Analysis prices Wonder at $4.52 per Intelligence Index task.
• Artificial Analysis Intelligence Index v4.3: Wonder 52 and Wonder Full 56, versus Muse Spark 1.3 at 48 and, per Latent Space, Claude Opus 5.5 at 58; Simon Willison says the index equates Wonder (Max) with GPT-6.1 Sol.
• Wonder Full: Coding Agents Index 68, tying Claude Code with Sonnet 5.5; 94 of 100 ProofBench problems; Terminal-Bench 4.0 from 33% to 57%; Meta-reported SWE-Marathon 57.5% versus Opus 5.5 at 46.3% and GPT-6 Astra at 42.5%. The Decoder…
• TechCrunch calls them Meta Superintelligence Lab’s first frontier models, trained at Prometheus in Ohio; Latent Space calls them a second-generation family about 14 months after the lab was founded and notes a same-day Muse Code quota bug…
• Access disagrees: TechCrunch, The Decoder, and Latent Space limit Wonder Full to Muse Code and keep it off the API, while The Verge says both models are available to Muse Code subscribers. Sources also differ on whether and when both will…
• Meta says the models are trained not to share private information without permission, to refuse harmful requests, and to stay usable for defensive code review; a fuller safety report was promised, and Latent Space says no detailed system…
• An Oct. 14 Verge column says Google’s Gemini 4 Argon launched with introductory pricing up to 80% below rivals; the Oct. 12 Verge piece had left any Muse comparison with Argon unclear. A MarkTechPost page contains only a headline and…