Meta Finally Joins the Frontier Race With Muse Wonder

Meta has spent the past year watching rivals set the pace. On October 12 it answered with Muse Wonder and a larger sibling, Muse Wonder Full, the first models from Mark Zuckerberg's Superintelligence Labs to post numbers that belong in the same conversation as the leading systems from Anthropic, OpenAI and Google.

On the closely watched Artificial Analysis Intelligence Index, Muse Wonder Full scored 56 and the standard Wonder scored 52, against 58 for Anthropic's Claude Opus 5.5. Reporting differs on how to describe that placement. VentureBeat notes that Wonder Full lands a point behind the top of the table and level with Anthropic's cheaper Sonnet 5.5. However you read the ranking, the scores themselves are consistent across outlets, and they put Meta back in frontier territory for the first time in a while.

There is a complication. The more capable Wonder Full is not being sold as an API product. Meta is reserving it for paying subscribers to Muse Code, its programming assistant, and offering only the weaker standard Wonder to developers who want to build on it. Pricing for that model runs to 4 dollars per million input tokens and 20 dollars per million output tokens, roughly in line with the going rate for a mid-tier frontier model.

Meta's claims, and the pushback

On Meta's own charts, Wonder Full beats OpenAI's GPT-6 Astra and Claude Opus 5.5 on several tests and does particularly well on a legal reasoning benchmark. Those are the company's figures, measured the company's way, and not everyone is buying them. Bindu Reddy, chief executive of Abacus.AI, wrote on X that the model was "nowhere near Fable or Astra level," with her own detailed results still to come. The gap between a vendor's slide and an independent run is exactly the sort of thing worth watching over the next few days.

What is not in doubt is Meta's appetite to spend. In a recent interview Zuckerberg said the company is prepared to put 100 to 200 billion dollars behind its AI effort, and framed the Muse family as a bet on agentic commerce rather than chatbots alone. That lands against a crowded month. Google shipped Gemini 4, OpenAI rolled out GPT-6 across ChatGPT, and Anthropic has been releasing a model roughly every couple of weeks, most recently the budget-tier Haiku 5.5.

The honest summary is that Meta has closed most of the distance without quite reaching the front. For a lab that many had written off a year ago, that is a real result. Whether the independent benchmarks hold up the headline, and whether gating the best model behind a coding subscription wins developers or frustrates them, is the part worth returning to.