Meta's Muse Wonder AI model launch

Meta's own benchmark claims and third-party verification tell different stories about where Muse Wonder actually lands. On the Artificial Analysis Intelligence Index v4.3, Muse Wonder Full scores 56 and the lighter Muse Wonder scores 52, compared with Claude Opus 5.5's 58, Claude Sonnet 5.5's 56, and GPT-6 Astra's 53 - tying Wonder Full with Sonnet 5.5 but leaving Opus 5.5 still ahead ^[1]. On the Artificial Analysis Coding Agents Index, Wonder Full reaches 68 at max effort, matching Claude Code on Sonnet 5.5, while Terminal-Bench 4.0 agentic performance jumped from 33% for the predecessor model to 57% for Wonder Full ^[1]. Wonder Full also solves 94 of 100 problems on ProofBench, a math-proof benchmark, up from 58 for its predecessor, and Meta reports its hallucination rate falling to 23% from 33% ^[1]. Meta's own launch materials frame the release as 'the next step on our scaling ladder and a major leap toward personal superintelligence' ^[2]. The same self-favorable framing shows up in Meta's own internal metrics: its 'Personal Superintelligence Index' claims Wonder Full scores 59.1, ahead of Claude Opus 5.5 and GPT-6 Astra, and its 'SWE Marathon' claim puts Wonder Full at 57.5 against 46.3 for Opus 5.5 - both benchmarks Meta created itself, in contrast with the independent Artificial Analysis Intelligence Index, which still puts Wonder Full at 56, behind Opus 5.5's 58. An independent coding benchmark tells a starker version of the same story: BridgeBench rated Muse Wonder at only 611 points - 8th of 29 models tested and 149 points behind the category leader ^[3]- and the benchmark's operator publicly accused Meta of gaming its self-reported scores. That gap between vendor-reported and independently verified results is the real story behind the launch.

Meta also split access in a way that limits how many people can test the stronger model directly. Muse Wonder is available through the Meta Model API and through Muse Code, Meta's subscription product, at $4 per million input tokens and $20 per million output tokens ^[4]. Muse Wonder Full, the larger and more capable of the two, is restricted entirely to Muse Code subscribers on the $50 and $200 monthly tiers, with no API access at all ^[4]. Muse Code itself is reported to run four tiers - $5, $15, $50, and $200 per month ^[5]. That structure concentrates Wonder Full's best capabilities inside Meta's own paid product rather than the open developer ecosystem, a notable departure for a company that built its AI reputation on giving models away.

The launch caps a strategic reversal that started well before October. Meta's AI effort began with the open-source Llama family, and by April 2025 Llama 4 had made Meta the leading open-source AI lab ^[6]. Zuckerberg was reportedly dissatisfied with Llama 4 and pushed a larger model, codenamed Behemoth, that was never released ^[6]. Meta Superintelligence Labs ^[7], under new chief AI officer Alexandr Wang, later shipped its first proprietary, closed model, Muse Spark, in April 2026 ^[8]- already a break from Meta's open-source identity ^[9]. Muse Spark 1.3 scored only 48 on the Artificial Analysis Intelligence Index in September 2026, the baseline that Muse Wonder and Wonder Full were built to beat ^[1], ahead of an October 2026 release that followed a reported 14-month development push ^[11]. Zuckerberg has backed that push with a reported capital expenditure buildout exceeding $100 billion, tied directly to the bet that frontier-level agentic models can close the gap with OpenAI and Anthropic ^[10].

Meta is pitching Muse Wonder less as a chatbot upgrade and more as a persistent agent. Active Memory is designed to retain conversation history, preferences, and habits across sessions, and Active Focus is meant to keep multi-step agentic tasks - the kind that hop between apps - from drifting off course; Meta says it put 'a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond' ^[2]. Industry analysts frame this as part of a broader shift from chatbots to agents enterprises can delegate real tasks to, but they are skeptical adoption will be fast: University of Washington professor Chirag Shah notes the industry is 'moving away from chatbots and going to this agentic world,' Gartner's Arun Chandrasekaran says Meta is 'really trying to steer these models to be more effective at personal agentic workflows,' and Tekonyx founder Sid Nag argues enterprise uptake depends on whether the functionality 'integrate[s] with existing applications and cybersecurity constructs' rather than on raw capability ^[4]. That trust question lands harder on Meta specifically than on other labs: online discussion around the launch repeatedly raised Meta's data-harvesting history, casting Active Memory - a feature built explicitly to remember personal details indefinitely - as a new surface for exactly the kind of data collection that has made Meta a reputational outlier among AI companies. Not every reaction has been this skeptical: some developers who tested the model directly on real coding tasks ranked it among the current top tier, pushing back on criticism - based on flashy demo benchmarks - that the model underperforms in realistic coding work.