Meta releases Muse Wonder and Wonder Full: intelligence index rises from 48 to 56

Meta has released Muse Wonder and Muse Wonder Full. On Artificial Analysis Intelligence Index v4.3, at maximum reasoning effort, the models score 52 and 56. Muse Spark 1.3, which came out just six weeks ago, scored 48. Wonder is available through Meta's own API and in Muse Code, while Wonder Full is so far available only to Muse Code users. The company positions both models for personal and professional agents.

Where the model improved

Muse Wonder Full improved on all ten index benchmarks and gained eight points. Agentic tasks rose the most. Terminal-Bench 4.0 climbed from 33% to 57%, GDPval-AA v2.1 from 1681 to 1857 Elo, and AA-Briefcase from 1581 to 1729 Elo. Full ranks in the top three on six of the ten benchmarks. It is first on CritPt (33%) and GDP.pdf (33%), but the margin is narrow: the closest competitors score 31–32%. On GDPval-AA it is second, and the result falls within the confidence intervals of Claude Opus 5.5 (1866) and Claude Sonnet 5.5 (1838). It places third on AA-Briefcase, AutomationBench-AA (71%) and SciCode (63%).

Muse Wonder gained four points. Its Terminal-Bench 4.0 score rose from 33% to 52%, and AutomationBench-AA from 58% to 68%. On the Artificial Analysis Coding Agents Index, Full in Muse Code scores 68, the same as Claude Code with Sonnet 5.5, while Claude Code with Opus 5.5 scores 66. Wonder scores 63, whereas Muse Code with Spark 1.3 scored 54.

Price: few tokens, but input is expensive

The models spend fewer output tokens than their predecessor: about 54 thousand per index task versus 60 thousand, or 10% fewer. That is lower than Gemini 4 Argon (62 thousand) and Claude Fable 5.1 (78 thousand), but higher than GPT-6.1 Sol (38 thousand) and GPT-6 Astra (27 thousand). Full also uses 16% fewer input tokens than Wonder: 6.44 million versus 7.69 million. It is both smarter and more economical.

The problem is the pricing. The rate card rose from $1.25 / $0.15 / $4.25 for Spark 1.3 to $4 per million input tokens, $0.20 for cached tokens and $20 for output tokens. Combined with an appetite for input context that is 28% larger than Spark 1.3's, this works out to $4.52 per index task, 2.8 times the previous $1.60. Input accounts for $3.43, or 76% of the cost: $1.92 for cache writes, $1.44 for cache hits and $0.07 for uncached input. Output makes up the remaining $1.08.

• Context: 1 million tokens, same as Spark 1.3
• Input modalities: text, image, video, audio; output is text only
• Wonder pricing: $4 input / $0.20 cache / $20 output per 1 million tokens
• Wonder Full is not sold by the token, so there is no per-task price
• Index task cost: cheaper than Claude Fable 5.1 ($7.63), Opus 5.5 ($5.98) and Sonnet 5.5 ($5.46)
• But more expensive than GPT-6 Astra ($3.26), Gemini 4 Argon ($1.99) and GPT-6.1 Sol ($0.72)

Knowledge and hallucinations

The weak spot of both models remains the same: factual knowledge. On AA-Omniscience, Full scores 30 and Wonder 25, on par with Spark 1.3. Claude Opus 5.5 scores 46, GPT-6 Astra 43, Gemini 4 Argon 42 and Claude Sonnet 5.5 32. A low hallucination rate softens the picture. Full's rate fell from 33% to 23%, the second-best result among models with an index of 45 or higher, behind Gemini 4 Argon (15%). Wonder, at 27%, ranks third. The model is more likely to say it doesn't know the answer than to make one up.

Who it suits and when

Wonder is worth trying where agentic scenarios matter: terminal tasks, workflow automation, documents and scientific calculations. Before moving from Spark 1.3, you need to redo your budget. In long agentic loops, input eats up three quarters of the cost, so the cache hit rate directly determines the bill. If the task depends on reference knowledge, look at Opus 5.5 or GPT-6 Astra instead. If cost per task is what matters most, GPT-6.1 Sol and Gemini 4 Argon are noticeably cheaper. Full is currently available only inside Muse Code and has no per-token price, so you cannot build it into your own pipeline. Keep that limitation in mind when planning.