Meta's Muse Wonder Full breaks into AI's top tier

Meta re-enters the frontier race

Meta launched two new frontier models on October 12, 2026: Muse Wonder and its more powerful sibling, Muse Wonder Full. Both land among the best AI models available today, marking a sharp comeback for a company that has struggled to keep pace with Anthropic and OpenAI over the past two years.¹

Muse Wonder Full scores 56 points on the Artificial Analysis Intelligence Index, while Muse Wonder reaches 52. That is a big jump from 48 for their predecessor, Muse Spark 1.3, which sat in the middle of the pack.¹ On the Vals AI Index, Muse Wonder Full debuts at fifth place and beats every OpenAI model.⁵[X]

Agentic coding leads the gains

The biggest improvements show up in agentic evaluations, which measure how well a model can handle complex tasks on its own. On Terminal-Bench 4.0, performance doubled from 33 to 57 percent.¹ Muse Wonder Full also topped the coding challenge index at 68 points, tying it with Claude Opus 5.5.

On ProofBench, a benchmark for math proofs, Muse Wonder Full solved 94 out of 100 problems, just a few points shy of the best models.¹

Both models support a one-million-token context window and accept text, images, video, and audio. They also include Active Memory, which carries key information across conversations, and Active Focus, which helps the model stick with hard tasks and ask follow-up questions when something is unclear.¹²

Factual knowledge remains a weak spot

Despite the coding wins, factual knowledge still lags. On a knowledge test called AA-Omniscience, Muse Wonder Full scored only 30 points, well behind Claude Opus 5.5 at 46 and GPT-6 Astra at 43.¹ That gap suggests the Wonder series may be built on a smaller base model than its top competitors.

Muse Wonder Full did cut its hallucination rate down to 23 percent from 33 percent, the second-lowest among models scoring at least 45 points. Only Gemini 4 Argon performed better at 15 percent.¹

Steeper price for stronger results

Better performance comes at a steep cost. Muse Wonder runs about $4.52 per task on the Intelligence Index, nearly three times the price of Muse Spark 1.3. Token pricing sits at $4 per million input tokens and $20 per million output tokens, up sharply from the predecessor's $1.25 and $4.25.¹²

Muse Wonder Full is not available through the API at all. It is locked behind Muse Code subscriptions, which start at $50 a month for the Power plan and $200 a month for Maximum.²³ The model will also power the consumer Muse app for paying subscribers starting later this year.

Meta trained both models at its Prometheus data center at gigawatt scale.¹ Whether the jump in ability justifies the jump in price will depend on how builders and everyday users respond when real work runs through these models.³

Key facts

Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index, placing it among the top frontier models.
Muse Wonder Full costs nearly three times more per task than its predecessor, Muse Spark 1.3.
Muse Wonder Full is only available through Muse Code subscriptions, not the API.
Muse Wonder Full hits 68 on the Coding Agents Index, tying Claude Opus 5.5.
Factual knowledge remains weak, with Muse Wonder Full scoring only 30 on AA-Omniscience.

Seen on social

Meta just surpassed every OpenAI model on the Vals Index. Muse Wonder Full (Max) debuts at #5 on the Vals Index, up 6.4pts from its predecessor. pic.twitter.com/7YdW1N8Wwr

— Vals AI on X, see the post