🔍 Read the full analysis: Muse Wonder Full Enters The AI Frontier, While Muse Wonder Recharges on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Meta has released Muse Wonder and Muse Wonder Full, six weeks after Muse Spark 1.3. Artificial Analysis rates Wonder Full at 56 on its Intelligence Index, while API-available Wonder scores 52; the evaluator estimates a benchmark task with Wonder costs $4.52, compared with $0.72 for GPT-6.1 Sol at the same score.
Meta has released Muse Wonder and Muse Wonder Full, two proprietary AI models that Artificial Analysis evaluated on October 12. Wonder Full scored 56 on the evaluator’s Intelligence Index v4.3, while Wonder, the model available through Meta’s API, scored 52; Artificial Analysis estimated one benchmark task with Wonder at $4.52, compared with $0.72 for GPT-6.1 Sol, which scored the same.
The models share a one-million-token context window and accept text, images, audio and video as input, returning text. Their availability differs: Muse Wonder is offered through Meta’s API and Muse Code, the company’s coding agent. Meta lists it at $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Muse Wonder Full is available only to Muse Code users, with no per-token price published in the source material.
Artificial Analysis reported that Wonder Full improved on all 10 evaluations in its index compared with Muse Spark 1.3, rising from 48 to 56. It placed in the top three on six evaluations. Its Terminal-Bench 4.0 result was 57%, up from 33% for Spark 1.3; the benchmark evaluates models’ ability to complete work in a shell. On GDPval-AA, a test of economically useful knowledge work, Wonder Full scored 1,857 Elo, close to Claude Opus 5.5’s 1,866 and Claude Sonnet 5.5’s 1,838. The source says those results are within the models’ confidence intervals.
Wonder’s results were lower but improved on several measures versus Spark 1.3: it scored 52% on Terminal-Bench and 68% on AutomationBench-AA, up from 58%. Artificial Analysis gives Wonder a 27% hallucination rate, compared with 33% for Spark 1.3. These are evaluator findings, not guarantees about performance on every task or in every deployment.
Infographic 01 · Where Meta’s new pair lands
Muse Wonder Full ties for second. Muse Wonder ties GPT-6.1 Sol.
Artificial Analysis scores 27 frontier and open-weight configurations on ten evaluations. Meta’s two new models land near the top of the list. Six weeks earlier, Muse Spark 1.3 scored 48.
Intelligence Index score, 27 configurations
Higher is better. Reasoning effort in brackets, as Artificial Analysis tested it.
Availability: Muse Wonder runs on Meta’s API and in Muse Code. Muse Wonder Full is only available to Muse Code users and has no per-token price.
Source: Artificial Analysis, Muse Wonder analysis (12 Oct 2026) and model leaderboard, read 13 Oct 2026. Intelligence Index v4.3, ten evaluations.
Powered by Thorsten Meyer AI
Strong Scores, Uneven Access and Cost
The release puts Meta’s model quality and its commercial offer on different footing. Artificial Analysis’s score of 56 for Wonder Full places it near leading models in the evaluator’s comparison, but customers cannot purchase that model through a token-priced API based on the information provided. The more accessible Wonder scores four points lower and carries a relatively high estimated cost per index task.
Artificial Analysis calculated that one Intelligence Index task costs $4.52 with Muse Wonder. Its comparison puts GPT-6.1 Sol at $0.72 for a score of 52, Gemini 4 Argon at $1.99 for a score of 53, and Claude Sonnet 5.5 at $5.46 for a score of 56. These are evaluator estimates for its benchmark workload, not a universal price for every customer’s use. Still, the gap matters to organisations comparing model quality against expected API spending.
The release also gives coding teams a notable result, with an important caveat. In Artificial Analysis’s Coding Agents Index, Muse Code with Wonder Full scored 68, tying Claude Code with Sonnet 5.5. That measures the model and its surrounding coding-agent system together. It does not establish that Wonder Full would receive the same result inside Claude Code or that the model alone outperforms another configuration.
As an affiliate, we earn on qualifying purchases.
Six Weeks After Muse Spark
The release follows Muse Spark 1.3 by six weeks, according to the source report. On Artificial Analysis’s Intelligence Index v4.3, Spark scored 48, compared with 52 for Wonder and 56 for Wonder Full. The new results suggest substantial gains in the evaluator’s measured tasks, especially shell-based work, but they do not by themselves establish performance across all real-world uses.
The two new models are not interchangeable from a buyer’s perspective. Wonder is the API product, with published per-token rates. Wonder Full is restricted to Muse Code, and the source provides no per-token pricing for it. The evaluator’s task-cost figures are useful for comparing tested workloads, while actual bills can vary with prompts, model outputs, caching and usage patterns.
Artificial Analysis also reported differences in input-token use. It estimated that Wonder processes 7.69 million input tokens per index task, compared with 6.03 million for Spark 1.3; Wonder Full was estimated at 6.44 million. The source says input accounted for $3.43, or 76%, of Wonder’s $4.52 estimated task cost. This helps explain the benchmark bill, but does not establish what any particular customer will spend.
“The stronger model is the one you can’t buy by the token, and the model you can buy competes one tier lower.”
— Thorsten Meyer, ThorstenMeyerAI.com
AI model performance evaluation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Pricing and Results Outside Benchmarks
Wonder Full’s per-token price is not provided in the source material, and its availability is limited there to Muse Code. That leaves buyers unable to compare its direct API cost with Wonder or other models on a per-token basis. The source also does not specify whether Meta plans to broaden access or publish pricing.
The reported benchmark scores and task-cost estimates come from Artificial Analysis evaluations; the material does not establish how results will translate to a customer’s particular workload. The cost comparison reflects one index task and should not be read as a fixed charge for all uses. Confidence intervals also matter: the source says several GDPval-AA results are within each other’s intervals, so small score differences do not necessarily indicate a clear performance separation.
Although reported hallucination rates declined from Spark 1.3 to both new models, the source gives no assurance that errors will be absent in deployed systems. The figures also do not explain how rates vary by task or prompt. Teams considering consequential uses would need their own testing and review procedures.
As an affiliate, we earn on qualifying purchases.
Access, Pricing and Independent Testing
The immediate next developments to watch are whether Meta publishes pricing or broader access for Muse Wonder Full, and whether its API offering changes. The source material does not give a timetable for either. Buyers can currently compare Wonder’s listed token rates and the evaluator’s benchmark estimates, while Muse Code users have access to the stronger model under the described arrangement.
Further independent evaluations and customer testing will help show whether the gains reported on Terminal-Bench, GDPval-AA and other index tasks carry over to specific coding and knowledge-work needs. For organisations weighing adoption, the practical questions are how the model performs on their own tasks, how much input and output it consumes, and what checks are required for errors.
AI development and benchmarking software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Meta release?
Meta released Muse Wonder and Muse Wonder Full. Both are proprietary models with a one-million-token context window and text, image, audio and video input, according to the source.
How did the models score?
Artificial Analysis scored Wonder Full at 56 and API-available Wonder at 52 on its Intelligence Index v4.3. Muse Spark 1.3 scored 48 in the comparison.
How much does Muse Wonder cost?
Meta’s listed API rates are $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Artificial Analysis estimated a $4.52 cost for one of its Intelligence Index tasks; that is a benchmark-specific estimate, not a universal task price.
Can customers buy Muse Wonder Full through the API?
The source says Muse Wonder Full is available only to Muse Code users. It provides no per-token price for the model, so its cost per benchmark task cannot be calculated from the stated pricing.
Do the benchmark results guarantee reliable answers?
No. Artificial Analysis reported improved scores and lower hallucination rates than for Spark 1.3, but the source does not establish performance on every real-world task. Its figures support comparison, not a guarantee; verification remains relevant for work where errors carry significant costs.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
