It's Actually Good

Summary

Ben Davis reviews Meta’s Muse Wonder models, concluding that the full version is a genuinely strong and distinctive release, though it does not yet match the best models from Anthropic or OpenAI. He sees Meta as a credible third competitor, but says confusing naming, access options, pricing, and usage limits undermine the experience.

Model versions and access

Davis strongly advises against the less capable, incomplete Muse Wonder version, which he says costs about as much as Opus but performs substantially worse. The more capable Muse Wonder full is available through Muse Code on the $50 or $200 monthly plans, and through T3 Code, which he describes as a wrapper for the Muse CLI.

His practical advice is to use T3 Code rather than the CLI. He found the CLI’s approval workflow cumbersome and less effective than the automatic approval modes in other coding tools.

Limits and value

The full model can consume subscription limits quickly. Davis estimates that a $50 plan’s five-hour allowance represented roughly $332 in equivalent API usage per month—about 6.6 times the subscription price. However, he still exhausted the plan’s limits during testing and upgraded to $200. He also reports using 9% of a five-hour allowance in ten minutes while running three agents.

In practice, he found the limits much less generous than Claude’s subscription. He says the model’s token efficiency and overall value need improvement.

Capabilities and communication

The standout feature for Davis is the model’s communication style. He finds it clear, pleasant, and distinct from the familiar style of Claude models. It can explain its work through concise bullet points, useful next steps, screenshots, and well-organized updates. In his experience, it also coordinates effectively with other agents and models.

Coding and visual output

In a Svelte-from-scratch coding test, Davis places Muse Wonder full around the level of the model he calls “6.1 Sol”: capable, but not exceptional, with some shortcomings in code quality. It handled HTML-based timeline visualizations and refinements to an existing product interface well.

However, Davis does not consider it a strong out-of-the-box designer and says it lacks the impressive 3D-generation abilities seen in some competing models.

Benchmarks and overall standing

Davis cites Artificial Analysis’ index, where Muse Wonder full scores 56, behind Opus 5.5, while the less capable Muse Wonder scores 52. He cautions that benchmarks and viral 3D demonstrations do not necessarily reflect everyday usefulness.

His overall judgment is that Opus remains clearly ahead in result quality, while Muse Wonder’s originality and agent orchestration make it promising.

Outlook

Davis is impressed by how quickly Meta has become competitive, but says the company still needs to improve pricing, clarify product distinctions, broaden availability, and further develop the models.

Firecrawl sponsorship

The video includes a Firecrawl sponsorship segment about Alexandria, a search feature for agents. Davis describes it as providing access to more than 113 million indexed sources and 177 data providers, with examples including website-traffic data and information about Y Combinator companies.

Main speakers and sources

• Ben Davis — reviewer and primary speaker
• Alexander Wang — cited in a post about Meta’s model development
• Artificial Analysis — cited for benchmark and index scores
• Firecrawl — sponsor featured in the video