Why Muse Wonder Is Dividing AI Users

Meta presented Muse Wonder and its larger sibling, Wonder Full, as frontier competitors, with Alexander Wang saying Wonder was intended to compete with “the big boys.” The benchmark results were mixed: Wonder Full scored 79% on DeepSuite 1.1, nearly five points ahead of GPT-6 Astra and Opus 55, but trailed them by two points on Terminal Bench 4.0. Artificial Analysis gave Wonder Full an intelligence-index score of 56, tied for second and two points behind Opus 55; Wonder scored 52 and ranked seventh.

The clearest strength was agentic work. Wonder Full led Meta’s personal-superintelligence index by about four points over Astra and Opus, and ranked state of the art on Harvey Lab, Val’s finance agent, and Automation Bench, while sitting a few points behind frontier rivals on OS World. Artificial Analysis called it a material gain over Muse Spark 1.3, released six weeks earlier, but estimated a cost of $4.52 per task—almost three times Spark’s, about $1 more than GPT-6 Astra or GROC-47, and about $1.50 less than Opus 55.

User reports showed how little those scores settled the question of practical performance. Theo ranked Wonder Full in his S tier alongside Opus 55, and Charlie Holtz said it found a real security vulnerability that other agents missed and impressed him on design, visualizations, and CLI details. BoneGPT found it neck and neck with or better than Astra, Fable, and Opus on technical benchmarks. Clairvaux ranked Muse first on agentic tasks in a test of eight frontier models, but found it weak at front-end and design work.

Other users reported disappointing coding results. Max Blade said Wonder was “nowhere near” Opus 55 on a 3D lawnmower-simulator task, while Mahu Mohan said it performed poorly on real programming work in his agents. Everystandshipper reported a 52% score for Wonder on a senior-engineer benchmark, compared with 71% for GPT Sol.61; Riley Brown said he returned to Codex and Claude Code after five hours with Wonder. Theo cautioned that social-media-friendly 3D demos may not reflect real coding work and could steer labs toward the wrong use cases.

The distinction between Wonder and Wonder Full also fed the split in reactions. Adam Holter said Full looked good in his tests while Wonder did not, arguing that the shared name reflected badly on Full; Wang called that feedback fair. Some users also said usage ran down unusually fast: Arafat Shuvow reported that Wonder consumed his weekly limit too quickly, and Matt Silv said Full used 13% of his five-hour allowance in under 10 minutes on a $50-a-month plan. Wang said a bug might have miscounted usage, while acknowledging the damage to initial impressions.

Meta described the models as assistants built for agents acting on users’ behalf, with active memory to track priorities across conversations and active focus to persist through difficult tasks and recover from errors. Reports suggested promise on specific jobs but not dependable execution: Satya Malik said Muse handled immediate tasks such as booking travel yet repeatedly gave him a false meeting-time alarm, and Sam Lesson abandoned a hotel booking after Hilton’s website put it into a loop. Brycent said he uses different assistants for different tasks, though switching between them can carry a cost.

Mark Zuckerberg and Wang called the Wonder family state of the art for personal and professional agentic work, and Wang said Meta had reached the frontier and was already working on the next step up. The models were not described as pushing the frontier, and they appeared to have strong cyber guardrails. Zuckerberg declined to give a P-doom estimate, saying he was closer to viewing AI as a technology that can be engineered responsibly with normal guardrails. That stance contrasted with labs that treat advanced AI as potentially godlike, leaving the disagreement unresolved.