> ## Content Index
> Fetch the complete content index at: https://theainewsfiles.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Meta releases Muse Wonder and Muse Wonder Full, and says they lead on agent tasks
- URL: https://theainewsfiles.com/meta-muse-wonder-full-release-benchmarks/
- Published: 2026-10-13T06:29:53.000Z
- Updated: 2026-10-14T06:55:30.000Z
- Description: Meta says its new Muse Wonder models deliver state-of-the-art performance on agentic tasks. Artificial Analysis's index scores put both above Muse Spark 1.3, and Muse Wonder costs more per task than its predecessor.
- Author: J E Cowie
- Tags: News

Meta introduced two new models on 12 October 2026: Muse Wonder and Muse Wonder Full. Meta's own announcement calls them "our most capable models that deliver state-of-the-art performance on personal and professional agentic tasks". The independent benchmarking firm Artificial Analysis published its own results the same day. This article separates what Meta says, what is available now, and what Artificial Analysis measured.

## What Meta announced

Meta describes the pair as "a major leap toward personal superintelligence" and as the next step on its scaling ladder. Its post says both models were trained at its gigawatt-scale Prometheus data center. It also says the company began a ground-up overhaul of its AI efforts fourteen months ago.

Meta highlights two named capabilities. It describes Active Memory as "carrying important context about you across conversations and actively retrieving the details relevant to the task at hand". Active Focus, in Meta's words, "describes how Muse Wonder continues working toward a challenging goal: trying alternative approaches, working through blockers, and adapting to new information". Meta says the model asks for clarification when needed and checks in before taking important actions.

Meta says the models achieve state-of-the-art results across key benchmarks for personal and professional agents, memory and coding. Those are Meta's claims about its own models. Its post points to a separate report for the evaluation details.

## What is available now

- **Muse Wonder** is available today through the Meta Model API and Muse Code, according to Meta.
- **Muse Wonder Full** is available today for Muse Code subscribers. Artificial Analysis likewise reports that it is currently available only to Muse Code users.
- **Muse** is Meta's personal agent. Meta says Muse Wonder "will soon power experiences in Muse, and will roll out first to subscribers". That is a stated plan, not a release.

Meta also says it will use Muse Wonder in future to develop specialised models for everyone using Muse. On safety, Meta says Muse Wonder offers its strongest safeguards to date against chemical, biological, radiological and nuclear (CBRN) and cyber misuse. It calls Muse Wonder "our most aligned model to date" and says a full report with additional safety evaluations "will be available soon".

![A server hall inside a large data centre, with rows of racks](https://theainewsfiles.com/content/images/2026/10/shot2_a-server-hall-inside-a-large-data-centre.jpg)

Illustrative image of a data centre.

## What the independent numbers show

Artificial Analysis scores models on its Intelligence Index (version 4.3). At maximum reasoning effort it reports 56 for Muse Wonder Full and 52 for Muse Wonder. Both are up on Muse Spark 1.3 (max), which scored 48, and Artificial Analysis notes a release gap of only six weeks. Its headline calls the result "a material jump in intelligence over previous Meta models".

Artificial Analysis reports the following:

- **Agentic gains.** Muse Wonder Full improved on Muse Spark 1.3 across all ten Intelligence Index evaluations. Its largest gains were on Terminal-Bench 4.0, up from 33% to 57%, which brings it level with [Gemini 4 Argon](https://theainewsfiles.com/gemini-4-argon-fairwind-bloomberg-skepticism/) (high). It also gained 176 Elo on GDPval-AA v2.1 and 148 Elo on AA-Briefcase.
- **Rankings.** Muse Wonder Full places in the top three on six of the ten evaluations. It is first on CritPt (33%) and GDP.pdf (33%), and second on GDPval-AA v2.1 (1857 Elo). Artificial Analysis calls the leads on CritPt and GDP.pdf narrow, with the nearest competitors on 31–32%. On GDPval-AA v2.1, the result is within the confidence intervals of Claude Opus 5.5 (1866) and Claude Sonnet 5.5 (1838).
- **Coding.** On its Coding Agents Index at max effort in Muse Code, Muse Wonder Full scores 68, alongside Claude Code with Sonnet 5.5 at 68 and Opus 5.5 at 66\. Muse Wonder scores 63, against 54 for Muse Code with Muse Spark 1.3.
- **Context and modalities.** Each model has a 1M-token context window. Inputs are text, image, video and audio; output is text.

## The cost, and a weak spot

Artificial Analysis lists Muse Wonder's pricing at $4 per million input tokens, $0.20 cached and $20 per million output tokens, up from $1.25, $0.15 and $4.25 for Muse Spark 1.3\. Muse Wonder Full is not priced per token, so there is no cost-per-task figure for it.

For Muse Wonder (max), Artificial Analysis reports a cost of $4.52 per Intelligence Index task, 2.8 times Muse Spark 1.3's $1.60\. It attributes this to higher per-token prices and higher input token use (7.69M per task). That is lower than Claude Fable 5.1 ($7.63), Claude Opus 5.5 ($5.98) and Claude Sonnet 5.5 ($5.46). It is higher than GPT-6 Astra ($3.26), Gemini 4 Argon ($1.99) and GPT-6.1 Sol ($0.72).

Artificial Analysis also calls knowledge "a relative weakness" for both models, mitigated by very low hallucination rates. On AA-Omniscience, Muse Wonder Full scores 30 and Muse Wonder scores 25\. Both are below Claude Opus 5.5 (46), GPT-6 Astra (43), Gemini 4 Argon (42) and Claude Sonnet 5.5 (32). Muse Wonder Full's hallucination rate fell from 33% to 23%.

## How to read it

Two sets of claims sit side by side. Meta's "state-of-the-art" language comes from the company that built the models. Artificial Analysis's scores are a separate test. They support a clear step up from Muse Spark 1.3 and a strong showing on several agentic and coding measures. They do not show a clean win across the board: the lead on two evaluations is narrow, one result is within the confidence interval of two Claude models, and knowledge scores trail.

The gap between announced and shipped matters here too. Muse Wonder powering experiences in Muse is, by Meta's own account, still to come. What can be used today is the API and the Muse Code assistant.

## Sources

Artificial Analysis (2026) 'Muse Wonder and Muse Wonder Full: a material jump in intelligence over previous Meta models', Artificial Analysis, 12 October 2026, available at: [https://artificialanalysis.ai/articles/muse-wonder](https://artificialanalysis.ai/articles/muse-wonder?ref=theainewsfiles.com) (Accessed: 13 October 2026).

Meta (2026) 'Muse Wonder: A Frontier Model For Personal & Professional AI Agents', Meta AI Research, 12 October 2026, available at: [https://research.meta.ai/blog/muse-wonder](https://research.meta.ai/blog/muse-wonder?ref=theainewsfiles.com) (Accessed: 13 October 2026).

## Picture credits

- Top picture. Image: Aerial view of Meta (Facebook) headquarters campus, Menlo Park, by Pi.1415926535, [CC BY-SA 3.0](https://creativecommons.org/licenses/by-sa/3.0/?ref=theainewsfiles.com), via [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Aerial%5Fview%5Fof%5FFacebook%5Fcampus%5Fin%5FMenlo%5FPark,%5FSeptember%5F2019.JPG?ref=theainewsfiles.com)
- Illustrative image of a data centre. Image: San Antonio data center, by Robert Scoble, [CC BY 2.0](https://creativecommons.org/licenses/by/2.0/?ref=theainewsfiles.com) (Flickr) ([source](https://www.flickr.com/photos/35034363287@N01/2340202215?ref=theainewsfiles.com))