Overview
Muse Wonder is Meta new flagship model line, released on 2026/10/12 with two variants, Muse Wonder and Muse Wonder Full. Artificial Analysis measured them on Intelligence Index v4.3 at maximum reasoning effort: Muse Wonder scores 52 and Muse Wonder Full scores 56, against 48 for the previous generation Muse Spark 1.3 at maximum effort, with only six weeks between the two releases.
It is the second model from the group led by Alexandr Wang since he joined Meta, and the first Meta product positioned explicitly against the frontier labs. Meta states the purpose plainly: both models are designed for personal and professional agent use.
Availability splits across the two. Muse Wonder went live on Meta first-party API and in Muse Code on day one, while Muse Wonder Full is currently limited to Muse Code subscribers.
Key Features
- Two variants: Muse Wonder and Muse Wonder Full, scoring 52 and 56 on the Artificial Analysis Intelligence Index
- Top three on six of ten evaluations: First on CritPt at 33% and GDP.pdf at 33%, second on GDPval-AA v2.1 with 1857 Elo, third on AA-Briefcase with 1729 Elo, AutomationBench-AA at 71% and SciCode at 63%
- Large jump on Terminal-Bench 4.0: Rises from 33% on the previous generation to 57%, level with other leading models in the same band
- Coding agent capability: Muse Wonder Full scores 68 on the Artificial Analysis Coding Agents Index and Muse Wonder 63, against 54 for the previous generation
- 1M token context: A 1M token context window, unchanged from the previous generation
- Multimodal input: Accepts text, image, video and audio input, and produces text output
- Fewer output tokens: Around 54k output tokens per Intelligence Index task, roughly 10% below the previous generation
- Active Memory and Active Focus: Mechanisms Meta describes as reducing agent drift from the goal on long tasks
Use Cases
- Agentic work that needs long context and multimodal inputs
- Running terminal operations and knowledge work automation through the API at scale
- Repository-level changes and automated coding in Muse Code
- Scientific and engineering reasoning tasks
- Teams already on Muse Spark evaluating the upgrade inside the same API
Pros
- Intelligence Index moves from 48 to 56 in six weeks, a sizable generational step
- Top three on six of ten evaluations, first on two of them
- 1M context with text, image, video and audio input
- Around a tenth fewer output tokens per task than the previous generation
- Muse Wonder Full scores 68 on the Coding Agents Index, level with the leading coding agents
Pricing
Muse Wonder is priced on Meta first-party API at $4 per million input tokens, $0.20 cached input and $20 output. Artificial Analysis measures cost at $4.52 per Intelligence Index task at maximum reasoning effort. Muse Wonder Full is not priced per token and is currently limited to Muse Code subscribers. Using the models through the Muse app requires a $20 or $100 monthly subscription, while Muse Code subscriptions run at $50 or $200 per month. Exact pricing follows Meta official announcements.
Summary
Muse Wonder is Meta answer in the flagship model race, released on 2026/10/12. Muse Wonder and Muse Wonder Full score 52 and 56 on the Artificial Analysis Intelligence Index v4.3, a clear step up from Muse Spark 1.3 at 48 just six weeks earlier, with top-three placings on six of ten evaluations and first place on two of them. Its strengths sit in agentic and scientific work: Terminal-Bench 4.0 rises from 33% to 57%, and Muse Wonder Full reaches 68 on the Coding Agents Index. Context stays at 1M tokens, input covers text, image, video and audio, and output tokens per task fall by roughly a tenth. Pricing is the thing to watch, at $4 in and $20 out per million tokens, with the Full variant available only inside Muse Code. If you already work inside the Muse ecosystem, or you want a model that places top three on knowledge-work agent evaluations, this generation is worth testing. If cost per task is the first number you look at, cheaper options exist at similar intelligence levels.
Version History
- Meta releases Muse Wonder and Muse Wonder Full, scoring 52 and 56 on the Artificial Analysis Intelligence Index (2026-10-12): On October 12, 2026, Meta released its new flagship model line, Muse Wonder and Muse Wonder Full. Artificial Analysis measured them on Intelligence Index v4.3 at maximum reasoning effort: Muse Wonder scores 52 and Muse Wonder Full scores 56, against 48 for the previous generation Muse Spark 1.3 at maximum effort, with six weeks between the two releases. Muse Wonder Full places in the top three on six of ten evaluations, first on CritPt at 33% and GDP.pdf at 33%, second on GDPval-AA v2.1 with 1857 Elo, and third on AA-Briefcase with 1729 Elo, AutomationBench-AA at 71% and SciCode at 63%. Terminal-Bench 4.0 rises from 33% on the previous generation to 57%. Both models carry a 1M token context window, accept text, image, video and audio input and produce text output, and use around 54k output tokens per Intelligence Index task, roughly 10% fewer than the previous generation. Muse Wonder is available on Meta first-party API and in Muse Code from launch at $4 per million input tokens, $0.20 cached input and $20 output, while Muse Wonder Full is not priced per token and is limited to Muse Code subscribers. Meta also added Active Memory and Active Focus to both models to reduce agent drift from the goal.