🔍 Read the full analysis: Muse Wonder Full Joins The AI Frontier; Here’s What It Means 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 scored the Code-only Wonder Full at 56 on its Intelligence Index, while the API-available Wonder scored 52 and cost $4.52 per index task in the evaluation—far more than some similarly scoring rivals.
Meta has released Muse Wonder and Muse Wonder Full, two proprietary AI models that differ in both performance and availability. In Artificial Analysis’s evaluation published October 12, Wonder Full scored 56 on the Intelligence Index, while the API-available Wonder scored 52; the evaluation estimated a Wonder task cost of $4.52, raising questions about its price against similar-scoring competitors.
Artificial Analysis’s Intelligence Index v4.3 gave Muse Wonder Full a score of 56, eight points above Muse Spark 1.3’s 48. The evaluator reported that Wonder Full improved on all ten index evaluations and ranked among the top three on six. Only Claude Opus 5.5 at its maximum setting scored higher among the 27 configurations compared, with 58; Claude Sonnet 5.5 at maximum also scored 56. These are benchmark results, not guarantees of performance on every task.
The two models are reported to accept text, images, audio and video, return text, and have a one-million-token context window. Their access differs: Muse Wonder is available through Meta’s API and in Muse Code, while Wonder Full is available only to Muse Code users. Meta’s listed API rates for Wonder are $4 per million input tokens, $0.20 for cached input, and $20 per million output tokens. Wonder Full has no per-token price, so its cost per task cannot be calculated from the information provided.
Artificial Analysis reported notable gains on individual tests. Wonder Full scored 57% on Terminal-Bench 4.0, up from 33% for Spark 1.3, and 71% on AutomationBench-AA. Its GDPval-AA score was 1,857 Elo, between Claude Opus 5.5 at 1,866 and Sonnet 5.5 at 1,838; the evaluator said the results fall within the models’ confidence intervals. In the Coding Agents Index, Muse Code paired with Wonder Full scored 68, matching the reported Claude Code and Sonnet 5.5 configuration. That index evaluates the model together with its software harness.
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
Performance Meets API Pricing
The release gives businesses and developers two distinct choices, rather than one model available at different settings. Wonder Full posts the stronger benchmark result, but access is limited to Muse Code and its per-token cost is not disclosed. The model available through Meta’s API is easier to integrate into external workflows, but scored lower in the reported index.
Artificial Analysis estimated that one Intelligence Index task cost $4.52 with Muse Wonder. Its comparison put GPT-6.1 Sol at $0.72 for a score of 52, Gemini 4 Argon at $1.99 for 53, and Claude Sonnet 5.5 at $5.46 for 56. These figures describe the evaluator’s benchmark workload; they are not a universal price for every customer task. Still, they give prospective users a reason to test their own token use and output quality before choosing the API model at its listed rates.
Cost is tied to how much work a model does, not just its per-token rates. Artificial Analysis said Wonder used 7.69 million input tokens per index task, compared with 6.03 million for Spark 1.3, while producing about 54,000 output tokens. The source’s cost breakdown attributes $3.43, or 76%, of the $4.52 task estimate to input. Long prompts or agent workflows that repeatedly read large context may make this pattern especially relevant, though actual bills will depend on use.
As an affiliate, we earn on qualifying purchases.
From Spark to Two Muse Models
The release came six weeks after Muse Spark 1.3. Artificial Analysis scored Spark at 48 on the same Intelligence Index, compared with 52 for Wonder and 56 for Wonder Full. The source material says Wonder’s Terminal-Bench score was 52%, AutomationBench-AA was 68%, and its hallucination rate was 27%; it reports 23% for Wonder Full, down from 33% for Spark 1.3.
Those error figures need to be read alongside the task and evaluation method. In the same comparison, Gemini 4 Argon had a reported hallucination rate of 15%. Artificial Analysis’s AA-Omniscience scores were 25 for Wonder and 30 for Wonder Full, below several listed competitors. The measures suggest improvement over Meta’s previous model, but do not establish that the new models are reliable enough to use without checks in consequential workflows.
Meta’s launch also separates its strongest benchmark configuration from its API product. Wonder Full’s coding-agent result reflects Muse Code plus the model; it does not show how Wonder Full would perform inside a different coding tool. Similarly, benchmark scores and the evaluator’s task-cost estimate offer a comparison under specific test conditions, not a complete account of latency, reliability, deployment costs or performance on an individual company’s data.
“The bill is in the input, not the output.”
— Thorsten Meyer, ThorstenMeyerAI.com
AI performance evaluation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Access, Costs and Reliability
The supplied material does not include a statement from Meta explaining the pricing strategy, the limited availability of Wonder Full, or when it might become available through the API. It also gives no per-token price for Wonder Full, so its operating cost cannot be compared directly with the listed API models.
The benchmark report does not establish how the models will perform across customers’ real workloads, or how their cost and reliability change with different prompts and tools. The reported scores and token counts reflect Artificial Analysis’s evaluation setup. The source material also does not provide enough detail to independently assess every test’s methodology or reproduce the cost estimates.
Hallucination rates and benchmark scores are not a substitute for task-specific validation. Although the reported figures show a decline from Spark 1.3, Wonder Full’s 23% rate remains higher than the 15% reported for Gemini 4 Argon in this comparison. Whether that difference matters will depend on the application, the consequences of errors and the strength of human review.
As an affiliate, we earn on qualifying purchases.
What Buyers Should Test
For now, developers weighing Muse Wonder against alternatives can compare models on representative prompts, tracking quality, input-token consumption and total cost rather than relying on index scores alone. They should also check how well each model handles their own documents, tools and error-sensitive tasks.
More information from Meta could clarify whether Wonder Full will reach the API, how its cost will be measured, and what service terms apply to both models. Until those details are available, the clearest confirmed distinction is that Wonder is the API option, while the higher-scoring Wonder Full is limited to Muse Code. The next meaningful comparison will depend on further product details and independent testing under real-world workloads.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Meta release?
Meta released Muse Wonder and Muse Wonder Full, proprietary models with reported support for text, images, audio and video input and a one-million-token context window.
How did the models score?
Artificial Analysis gave Wonder Full a 56 and Wonder a 52 on its Intelligence Index v4.3. These are results from that evaluator’s benchmark, not assurances of performance on every task.
Can customers buy Muse Wonder Full through the API?
The supplied material says Wonder Full is available only in Muse Code. Muse Wonder is available through Meta’s API and in Muse Code. No per-token price is given for Wonder Full.
How much does a Muse Wonder benchmark task cost?
Artificial Analysis estimated $4.52 per Intelligence Index task for Muse Wonder in its evaluation. The estimate reflects that workload and should not be treated as a fixed cost for every customer use.
Does the reported improvement mean the models need no verification?
No. The source reports lower hallucination rates than for Muse Spark 1.3, but gives rates of 23% for Wonder Full and 27% for Wonder. Users should test outputs and keep appropriate checks in place, particularly for error-sensitive work.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
