Meta's New Muse Wonder Models Take Aim at GPT-6 and Claude

Meta is rolling out two new AI models, Muse Wonder and Muse Wonder Full, and the company isn't shy about who it thinks they beat. According to Meta, the models hold their own against some of the industry's most capable systems, including OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Claude Fable 5.1, based on internal testing across several benchmarks.

Both models go live today for subscribers of Muse Code, Meta's coding-focused subscription tier. Wonder is also available more broadly through the Meta Model API, giving developers outside that subscriber base a way to tap into it.

This is the second AI model launch from Meta in short order. It follows Muse Spark, the debut release from the company's restructured Superintelligence Labs division, and comes alongside Muse, Meta's push into AI agents that's already generated plenty of buzz. Wonder and Wonder Full slot into that same family, and Meta says the plan is to eventually have Wonder power the Muse agent itself, starting with paying subscribers before a wider rollout. The company also says it's building more specialized models that it wants to "bring to everyone using Muse." Meta hasn't said when any of that happens or offered further detail on what those specialized models will do.

One notable gap in Meta's benchmark comparisons: Google's Gemini 4 Argon is absent. That's likely less a snub than a timing issue. Gemini 4 Argon is still new enough that its real-world capabilities haven't fully settled into focus, making a direct comparison harder to draw.

The bigger story here is Meta's pace. After a stretch where the company looked like it was losing ground to OpenAI, Anthropic, and Google in the foundation-model race, it's now pushing out models and products at a clip that suggests real urgency. Meta has committed billions of dollars to AI talent and the infrastructure needed to train and run these systems, and the Muse lineup has grown quickly to include not just Wonder and Wonder Full but also Muse Image, an image-generation model, alongside the Muse agent itself.

The through-line is agents. Meta's leadership has made clear it wants AI that does things for people, not just answers questions, and Muse is the vehicle for that bet. Having stronger underlying models like Wonder matters because an agent is only as capable as the model reasoning behind it. If Meta can make good on its claim that Wonder competes with the field's best, that's a meaningful signal that the company's AI reset is producing results rather than just headlines.

Whether Wonder and Wonder Full actually hold up against GPT-6 Astra and Claude's latest versions in practice, rather than in Meta's own benchmark tests, is the open question. Internal testing tends to favor the company running it, and independent evaluation will matter more once these models reach a wider audience. For now, Meta has staked out its claim. The next move is making Wonder the engine behind Muse for everyone, not just subscribers, and that's the rollout worth watching.