Reviews · OCTOBER 14, 2026
Meta's Muse Wonder lands at the frontier, but the strongest variant isn't on the API
Meta Superintelligence Labs shipped Muse Wonder and Muse Wonder Full on October 12, pricing the API model at $4/$20 per million tokens — while reserving Wonder Full, which leads DeepSWE 1.1 at 79.0%, for Muse Code subscribers at $50–$200 per month.
Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, pricing the API-accessible model at $4 per million input tokens and $20 per million output tokens, and reserving the stronger Wonder Full behind Muse Code, a developer subscription that runs $50 to $200 per month. For small teams building agent stacks, that pricing structure is the real story: Meta has entered the frontier tier at Opus 5.5's price point, but the two variants aren't interchangeable.
On Artificial Analysis's Intelligence Index v4.3, Wonder Full scores 56, matching Claude Sonnet 5.5. The API-exposed Wonder scores 52. GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1 all land at 53. Claude Opus 5.5 still leads the index at 58. So the model you can actually call over HTTP sits a notch below the current GPT and Gemini flagships; the model that beats them is behind a subscription.
The coding story is where Wonder Full earns its tier. It posts 79.0 on DeepSWE 1.1, ahead of Opus 5.5 at 74.2 and GPT-6 Astra at 74.1. On SWE-Marathon, Wonder Full hits 57.5 versus 46.3 for Opus 5.5. Its Coding Agents Index score of 68 ties Claude Code running on Sonnet 5.5. Terminal-Bench 4.0 is the exception: 57.1 against 59.1 for Astra and 59.6 for Opus 5.5. For context on where terminal performance has been trending, see Claude Sonnet 5.5's jump to 70.6 on Terminal-Bench 4.0.
Knowledge breadth is the weak spot. Wonder Full scores 30 on AA-Omniscience and Wonder scores 25, well behind Opus 5.5 at 46, Astra at 43, and Argon at 42. Hallucination rate on Wonder Full came in at 23%, down from a baseline of 33%, which Artificial Analysis flags as the second-lowest among models above 45 index points. Both variants carry roughly a one-million-token context window, and the Active Memory feature retains user context across days and weeks.
For a builder weighing routing decisions, the arithmetic is specific. Wonder's API price matches Opus 5.5 but trails it on the index. Compare that to GPT-6.1 Sol's one-fifth-of-Astra pricing or Gemini 4 Argon's AutomationBench lead, and the economics look crowded. Wonder Full is the interesting product, and it isn't on the API.
That split is the decision. Meta has shipped a credible frontier model and gated the version worth integrating behind a seat license, which is the kind of packaging choice that reveals how the vendor expects developers to actually use it.
Sources
- https://techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder/
- https://www.cnbc.com/2026/10/12/meta-muse-wonder-ai-model-agent.html
- https://the-decoder.com/metas-new-frontier-model-muse-wonder-full-breaks-into-the-ai-top-tier/
- https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
- https://venturebeat.com/technology/metas-new-muse-wonder-and-muse-wonder-full-models-achieve-frontier-performance-but-only-the-weaker-one-is-available-via-api