What Does Muse Wonder Full Bring To The AI Frontier?

🔍 Read the full analysis: What Does Muse Wonder Full Bring To The AI Frontier? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Meta has released Muse Wonder and Muse Wonder Full, two proprietary AI models that accept text, images, audio and video, following Muse Spark 1.3 by about six weeks. In an evaluation published October 12 by Artificial Analysis, Wonder Full scored 56 on the Intelligence Index v4.3, while the API-available Wonder scored 52; the results place Meta’s models near leading systems, but the reported cost of running Wonder is substantially higher than that of some models with similar scores.

Artificial Analysis reported that Muse Wonder Full scored 56, eight points above Muse Spark 1.3’s 48, and improved on all ten evaluations in the index. Among 27 configurations compared in the source report, only Claude Opus 5.5 at its maximum setting scored higher, with 58; Claude Sonnet 5.5 at its maximum setting also scored 56. The report cautions that benchmark scores are measurements under particular evaluation conditions, not guarantees of performance in every use.

The models differ in access. Muse Wonder is available through Meta’s API and Muse Code, its coding agent. Its listed price is $4 per million input tokens and $20 per million output tokens, with cached input listed at $0.20 per million tokens. Wonder Full is available only to Muse Code users, and the source report says it has no per-token price, leaving its cost per task unknown.

Artificial Analysis put Muse Wonder Full at 57% on Terminal-Bench 4.0, compared with 33% for Muse Spark 1.3, and reported a score of 1,857 Elo on GDPval-AA, a benchmark of economically useful knowledge work. The API model scored 52 on Terminal-Bench and 68% on AutomationBench-AA, up from 58% for Spark. These are evaluation results reported by Artificial Analysis, not independent evidence of how the models will perform across all real-world tasks.

Strong Scores, Uneven Access and Cost

The release suggests Meta has narrowed the measured performance gap with leading AI systems, particularly for coding and knowledge-work evaluations. Artificial Analysis’s Coding Agents Index scored Muse Code with Wonder Full at 68, tying its Claude Code and Sonnet 5.5 configuration. The index measures a model together with its software harness; it does not establish that Wonder Full would perform the same way inside Claude Code or that the models are interchangeable.

For customers, the practical divide is between the stronger model and the purchasable API model. Wonder Full’s results may matter to users already working in Muse Code, but the lack of a published per-token price makes its running cost difficult to compare. Wonder’s API pricing is public, yet the source’s benchmark-based estimate puts one Intelligence Index task at $4.52. GPT-6.1 Sol scored the same 52 for a reported $0.72 per task, while Gemini 4 Argon scored 53 for $1.99. Those figures are specific to the cited evaluation and should not be treated as universal costs for other workloads.

That gap matters to developers deciding whether benchmark gains justify operating expense. The source attributes much of Wonder’s estimated bill to input: it says the model processed 7.69 million input tokens per task, accounting for $3.43, or 76% of the $4.52 estimate. The evaluation does not establish why the model used that many tokens or whether customers’ applications would have the same usage pattern.

From Spark to Two Wonder Models

The release follows Muse Spark 1.3 by about six weeks, according to the source report. Artificial Analysis scored Spark 1.3 at 48 on the Intelligence Index; Wonder Full’s reported score of 56 and Wonder’s score of 52 mark gains in that evaluation. The source says Wonder costs $2.8 times as much per index task as Spark, comparing its estimate of $4.52 with $1.60 for its predecessor.

Both new models are described as having a one-million-token context window, accepting text, images, speech or audio, and video as input, and returning text. The report lists Wonder’s API prices at the same level as Claude Opus 5.5: $4 per million input tokens and $20 per million output tokens. It lists several other compared models, including Claude Sonnet 5.5, Gemini 4 Argon and GPT-6.1 Sol, at $2 per million input and $10 per million output tokens.

Benchmark results also show areas where Meta’s models remain behind. Artificial Analysis gave Wonder Full an AA-Omniscience score of 30 and Wonder a 25, compared with 46 for Claude Opus 5.5, 43 for GPT-6 Astra and 42 for Gemini 4 Argon. The source reports hallucination rates of 23% for Wonder Full and 27% for Wonder, down from 33% for Spark 1.3, but says Gemini 4 Argon recorded 15% among the models it compared with Intelligence Index scores of at least 45. The figures support caution in workflows where an incorrect answer carries costs.

Costs and Performance Outside Benchmarks

The figures available here do not establish what Muse Wonder Full costs per task, because the model is limited to Muse Code and has no per-token price in the source report. They also do not show whether the benchmark’s estimated API task costs match typical customer workloads, which can vary with prompts, context and usage.

The source material does not provide enough methodological detail to assess every benchmark’s uncertainty, reproduce the cost calculations, or establish how results will hold across different applications. The Coding Agents Index compares model-and-harness combinations, so its tie between Muse Code with Wonder Full and Claude Code with Sonnet 5.5 should not be read as a direct model-only comparison. Meta’s explanation for the difference between Wonder and Wonder Full, and whether wider access to Full is planned, is not included.

Pricing and Wider Availability

The next practical questions are whether Meta will make Muse Wonder Full available beyond Muse Code and whether it will publish a pricing basis that customers can compare with API models. For developers considering Muse Wonder, the reported input-token use makes it worth measuring cost on their own prompts and tasks rather than relying on a single benchmark estimate.

Further independent evaluations could show whether the reported gains persist across tasks and testing methods, while additional information from Meta could clarify model differences, usage costs and access plans. The source material does not identify a date for any such release, pricing update or follow-up evaluation.

Key Questions

What are Muse Wonder and Muse Wonder Full?

They are two proprietary Meta AI models. The source says both accept text, images, audio and video, have a one-million-token context window, and return text.

How did the models score in the cited evaluation?

Artificial Analysis reported scores of 56 for Muse Wonder Full and 52 for Muse Wonder on its Intelligence Index v4.3. The report also describes gains for Wonder Full across all ten evaluations in that index.

Can developers access Muse Wonder Full through Meta’s API?

Not according to the source report. It says Muse Wonder is available through Meta’s API, while Wonder Full is available only to users of Muse Code.

Why is Muse Wonder’s reported benchmark cost a concern?

The report estimates $4.52 per Intelligence Index task for Wonder, compared with $0.72 for GPT-6.1 Sol, which received the same index score of 52. These are evaluation-specific estimates, not guaranteed prices for every customer workload.

What remains unknown about the release?

The source does not give a per-token or per-task price for Muse Wonder Full, explain whether it will become available through the API, or provide a timeline for wider access. Real-world costs and performance may also differ from the cited benchmark results.

Source: ThorstenMeyerAI.com