Muse Wonder and Muse Wonder Full Mark a Significant Gain Over Earlier Meta Models

Muse Wonder and Muse Wonder Full Mark a Significant Gain Over Earlier Meta Models

Meta has released Muse Wonder and Muse Wonder Full, two models that score 52 and 56, respectively, on the Artificial Analysis Intelligence Index v4.3 at maximum reasoning effort. Both improve on Muse Spark 1.3 at maximum effort, which scored 48. The two models were released only six weeks apart from their predecessor.

Muse Wonder is available through Meta's first-party API and Muse Code. Muse Wonder Full is currently limited to Muse Code users. Meta describes both models as intended for personal and professional agent use.

Key results

• Muse Wonder Full shows the larger overall improvement. At maximum effort, it gains 8 Intelligence Index points over Muse Spark 1.3, improving across all ten evaluations. Its largest gains appear in agentic evaluations: Terminal-Bench 4.0 rises from 33% to 57%, GDPval-AA v2.1 increases from 1681 to 1857 Elo, and AA-Briefcase increases from 1581 to 1729 Elo.
• Muse Wonder reaches a score of 52. It gains 4 points over Muse Spark 1.3, with notable improvements on Terminal-Bench 4.0, from 33% to 52%, and AutomationBench-AA, from 58% to 68%.
• Muse Wonder Full ranks in the top three on six of ten evaluations. It is particularly strong in agentic knowledge work and scientific reasoning. It ranks first on GDP.pdf at 33%, second on GDPval-AA v2.1 at 1857 Elo, and third on AA-Briefcase at 1729 Elo and AutomationBench-AA at 71%. In scientific reasoning, it ranks first on CritPt at 33% and third on SciCode at 63%.
• Both models use fewer output tokens than Muse Spark 1.3. Muse Wonder and Muse Wonder Full each use about 54,000 output tokens per Intelligence Index task, approximately 10% fewer than Muse Spark 1.3 at 60,000. This is below Gemini 4 Argon at high effort, with 62,000 tokens, and Claude Fable 5.1 at maximum effort with fallback, with 78,000 tokens. It is above GPT-6.1 Sol at maximum effort, with 38,000, and GPT-6 Astra at maximum effort, with 27,000.
• Muse Wonder Full is more token-efficient than Muse Wonder. It uses 16% fewer input tokens per task, 6.44 million compared with 7.69 million, while achieving the higher Intelligence Index score.
• Muse Wonder's cost per task is higher than that of most similarly capable models. Its premium API pricing and high input-token use produce a cost of $4.52 per Intelligence Index task. This is below Claude Fable 5.1 at $7.63, Claude Opus 5.5 at $5.98, and Claude Sonnet 5.5 at $5.46. It is above GPT-6 Astra at $3.26, Gemini 4 Argon at $1.99, GPT-6.1 Sol at $0.72, and Muse Spark 1.3 at $1.60. Muse Wonder Full is not priced per token, so a comparable cost-per-task figure is unavailable.
• Both models improve coding-agent performance. Muse Wonder Full scores 68 on the Artificial Analysis Coding Agents Index at maximum effort in Muse Code, matching Claude Code with Sonnet 5.5 at 68 and exceeding Opus 5.5 at 66. Muse Wonder scores 63. Muse Code with Muse Spark 1.3 scores 54.

Model details

• Context window: 1 million tokens for each model, unchanged from Muse Spark 1.3
• Muse Wonder pricing: $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens
• Muse Wonder Full pricing: Not priced per token
• Input modalities: Text, image, video, and audio
• Output modality: Text
• Availability: Muse Wonder is available through Meta's first-party API and Muse Code. Muse Wonder Full is available only to Muse Code users.

Muse Wonder Full evaluation results

At maximum effort, Muse Wonder Full ranks in the top three on six of the ten Artificial Analysis Intelligence Index v4.3 evaluations:

• First: CritPt, 33%, and GDP.pdf, 33%
• Second: GDPval-AA v2.1, 1857 Elo
• Third: AA-Briefcase, 1729 Elo; AutomationBench-AA, 71%; and SciCode, 63%

The leads on CritPt and GDP.pdf are narrow. The nearest competing models score between 31% and 32%. On GDPval-AA v2.1, Muse Wonder Full's result falls within the confidence intervals of Claude Opus 5.5 at maximum effort with fallback, which scores 1866, and Claude Sonnet 5.5 under the same conditions, which scores 1838.

Its largest improvement over Muse Spark 1.3 is on Terminal-Bench 4.0, where the score rises from 33% to 57%, matching Gemini 4 Argon at high effort.

Pricing and token use

Muse Wonder costs $4.52 per Artificial Analysis Intelligence Index task, 2.8 times the $1.60 cost of Muse Spark 1.3. The increase reflects both higher token prices and greater input-token use.

Token pricing

Muse Wonder is priced at $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens. Muse Spark 1.3 costs $1.25, $0.15, and $4.25, respectively.

Muse Wonder's input price matches Claude Opus 5.5. It is twice the $2 per million input and $10 per million output prices listed for Claude Sonnet 5.5, Gemini 4 Argon, and GPT-6.1 Sol. It is 60% below the $10 per million input and $50 per million output prices listed for GPT-6 Astra and Claude Fable 5.1.

Input tokens

Muse Wonder uses 7.69 million input tokens per task, 28% more than Muse Spark 1.3 at 6.03 million. Its input use is between two and seven times that of Gemini 4 Argon, GPT-6.1 Sol, and GPT-6 Astra.

Input processing accounts for $3.43, or 76%, of Muse Wonder's cost per task:

• $1.92 from cache writes
• $1.44 from cache hits
• $0.07 from uncached input

Output tokens

Muse Wonder uses about 54,000 output tokens per task, costing approximately $1.08. This is 10% fewer output tokens than Muse Spark 1.3, which uses about 60,000 per task.

Knowledge and hallucination rates

Knowledge remains a relative weakness for both models. On AA-Omniscience, Muse Wonder Full scores 30 and Muse Wonder scores 25, matching Muse Spark 1.3. These results are below Claude Opus 5.5 at 46, GPT-6 Astra at 43, Gemini 4 Argon at 42, and Claude Sonnet 5.5 at 32.

Muse Wonder Full's 5-point gain over Muse Spark 1.3 comes from reducing its hallucination rate from 33% to 23%. Among models scoring at least 45 on the Intelligence Index, this is the second-lowest hallucination rate, behind Gemini 4 Argon at high effort, which records 15%. Muse Wonder records a 27% hallucination rate and ranks third by this measure.

Summary

Muse Wonder and Muse Wonder Full deliver substantial gains over Muse Spark 1.3 across general intelligence, agentic work, scientific reasoning, and coding-agent evaluations. Muse Wonder Full reaches an Intelligence Index score of 56 and places in the top three on six evaluations. Muse Wonder has a published API price and costs $4.52 per Intelligence Index task, while Muse Wonder Full is currently available through Muse Code without a per-token price.