Tech • AI • Robotics • Game
Meta Drops Muse Wonder Full: Beats GPT-6 Astra and Gemini 4 Argon
9.4/10TL;DR
Meta has launched Muse Wonder and Muse Wonderful, multimodal AI models aimed at long, autonomous coding and office tasks, with benchmark gains that place them near or above top rivals in several agent and software tests.
KEY POINTS
New models target long autonomous work
Meta Super Intelligence Labs released Muse Wonder and the more powerful Muse Wonderful after 14 months of development. Both support 1 million tokens of context, accept text, images, video and audio, and return text responses. Unlike the smaller Muse Spark, they are positioned for sustained work in coding, agents, law and finance.
Meta’s own benchmarks show a coding leap
On Meta’s internal Personal Super Intelligence Index, Wonderful scored 59.1%, ahead of Opus 5.5 at 55.2% and Astra at 54.5%. In Deep Suite, a software benchmark spanning 113 tasks, Wonderful posted 79% and Wonder 78.8%, versus roughly 74% for Opus and Astra. In the longer SWE Marathon test, Wonderful reached 57.5%, above Opus at 46.3% and Astra at 42.5%.
Independent evaluators found a real improvement
Artificial Analysis called the release a material jump. On its intelligence index, Wonderful scored 56, tying Claude Sonnet 5.5, beating several rivals around 53, and trailing only Opus 5.5 at 58. Spark 1.3 had scored 48 six weeks earlier, highlighting the size of the gain.
Performance is strongest in agents, weaker in raw knowledge
On Terminal Bench 4.0, which measures file handling and command-line autonomy, the new model rose from 33% to 57%. On legal work, Wonderful scored 13.1 on Harvey’s benchmark versus 4.2 for Opus. But on a 6,000-question knowledge test, Wonderful scored only 30, well below Opus at 46 and Astra at 43.
Hallucination rates improved significantly
When the model did not know an answer, it produced fabricated responses 23% of the time, down from 33% for Spark. That compares with 51% for Astra and 59% for Opus. Only Argon performed better on that measure at 15%.
Developers reported faster coding and useful catches
Early users said the model found practical issues in real codebases. Charlie Holtz of Conductor said Wonder identified a genuine security flaw that other agents had missed. Another developer reported Muse Code with Standard Wonder built a comparable 3D game in 3 minutes 42 seconds, versus 11 minutes 40 seconds for Claude Code running Opus.
Benchmark caveats remain substantial
Several leading scores came from tests designed by Meta, and the company acknowledged that rival models may not have been run with ideal settings or tools. On Program Bench, which focuses on building software from scratch, Wonderful scored 59, ahead of Astra at 52.5 but far behind Opus at 70.5. Some rankings were also described as within the margin of error.
Pricing and access are uneven
Standard Wonder costs $4 per million input tokens and $20 per million output tokens, matching Opus 5.5 pricing. Wonderful is not broadly available through the API and is limited to Muse Code subscribers on $50 and $200 monthly plans. Cost per task remains sensitive to the model’s heavy reading behavior, with average input usage of about 7.69 million tokens.
Safety review flagged residual bio-chemical concern
Meta said testing could not fully rule out high-risk potential in unrestricted chemical and biological use. After safety tuning and system filters, remaining risk was rated moderate or lower across categories. The company also stressed avoiding benchmaxing, or overtraining on famous tests, while arguing that larger base models generalize better to unfamiliar tasks.
The release reflects a broader reorganization at Meta
The push follows a turbulent period in which Mark Zuckerberg spent $14.3 billion in a deal tied to Scale AI founder Alexander Wang, who became chief AI officer in June 2025. Wang, 29, has said he is not an AI researcher, a point critics including Yann LeCun have raised. Meta also cut about 600 AI staff tied to its older research structure while recruiting aggressively, reportedly with signing bonuses of up to $100 million.
CONCLUSION
Meta’s new models appear to have closed much of the gap in coding and agent performance, but the strongest claims still depend partly on house benchmarks and limited access. The larger strategic test will be whether those gains hold up in broad real-world use across Meta’s consumer and enterprise products.
Ask a question