Meta has released two new AI models—Muse Wonder and Muse Wonder Full—positioning them as some of the company’s most capable systems yet and framing them as a step toward what it calls “personal superintelligence.” The announcement centers on two agent-oriented features: Active Memory and Active Focus, which are designed to help AI systems carry context across interactions and remain effective on complex tasks.
Meta also shared pricing and access details, while independent evaluation data from Artificial Analysis suggests strong reasoning performance—though Meta’s claims were met with skepticism from at least one prominent model evaluator. For builders and investors, the key question is not just raw benchmark placement, but how these “active” capabilities translate into real-world reliability and cost over time.
Key takeaways
- Meta introduced Muse Wonder and Muse Wonder Full, adding two agent features: Active Memory and Active Focus.
- According to Artificial Analysis, Wonder Full scored 56 and standard Wonder scored 52 on its Intelligence Index at maximum reasoning effort.
- Artificial Analysis’ results place Wonder Full ahead of Claude Fable 5.1 and GPT-6 Astra, but below Claude Opus 5.5.
- Meta’s pricing for the standard Wonder model is $4 per million input tokens, while Wonder Full is restricted to paid Muse Code tiers.
- Abacus.AI CEO Bindu Reddy publicly disputed the magnitude of Meta’s performance claims, signaling further scrutiny from independent benchmarks.
Meta’s “active” capabilities: memory and sustained attention
In a post on Meta’s research blog, the company said Muse Wonder and Muse Wonder Full add two capabilities it views as essential for more capable agents. Active Memory is intended to carry relevant information across conversations, rather than limiting the system to the immediate prompt window. Active Focus, Meta said, helps an agent keep working on difficult tasks—an approach aimed at reducing the tendency for models to lose track or underperform as tasks become more complex.
For users, these features matter because “better answers” aren’t only about one-off performance. If a system can recall what’s relevant from earlier exchanges and maintain attention on hard problems, it may require fewer manual interventions—such as restating requirements or reassembling context—which is a practical constraint for businesses evaluating AI agents.
How the models compare in independent testing
Independent evaluation data was a major part of the initial reception. Artificial Analysis, an evaluation platform, published scoring for Muse Wonder at 52 and Muse Wonder Full at 56 on its Intelligence Index, both measured at maximum reasoning effort.
On that same basis, Artificial Analysis’ reported comparisons show Wonder Full above Claude Fable 5.1 and GPT-6 Astra (each at 53), but behind Claude Opus 5.5, which scored 58. In other words, Meta’s top-tier option appears competitive but not the clear leader in this particular framework.
Still, benchmark outcomes should be read carefully. Different systems can excel under different evaluation setups—particularly when one model is framed as “agentic” and another is optimized around prompt-based tasks. Artificial Analysis’ numbers are useful as a snapshot, but they don’t automatically indicate how each model will behave over long, tool-using workflows.
Cost and access: where Meta draws the line for “Full”
Meta positioned Muse Wonder as its most affordable option among the two, describing it as its most expensive model to date. The standard Wonder model is priced at $4 per million input tokens. Meta said this pricing matches Claude Opus 5.5 at the same rate, while undercutting GPT-6 Astra at $10 and Claude Fable 5.1 at $50, based on the figures cited in the announcement.
Wonder Full, however, is not broadly available on a pay-as-you-go basis. Meta said access is limited to people subscribed to Muse Code’s $50 per month “power usage” tier or $200 per month “maximum usage” tier. That structure suggests Meta is treating the “Full” model as a higher-demand, more resource-intensive product—an approach that can shape early adoption, tooling integration timelines, and customer sentiment.
For teams planning deployments, access constraints can matter as much as model quality. If the best-performing model is tied to certain tiers, then cost predictability and availability may be the deciding factors when selecting an architecture.
Pushback from independent evaluators
Not everyone accepted Meta’s performance positioning without caveats. Abacus.AI CEO Bindu Reddy publicly challenged the company’s claims on X, writing that Muse Wonder was “nowhere near Fable or Astra level.” Reddy also said detailed benchmark results would follow, pointing to an expectation of further scrutiny beyond Meta’s own framing.
That kind of disagreement is common in the fast-moving model landscape, but it raises a practical takeaway for readers: when benchmark claims circulate, investors and builders should watch for follow-up evaluations and methodology transparency. Meta’s launch adds momentum, but independent cross-checking can determine whether the market’s interpretation aligns with the evidence.
What to watch next
As Muse Wonder and Muse Wonder Full roll out, the next signals to track are how Active Memory and Active Focus perform outside controlled benchmarks—especially in longer, agent-style workflows—and whether independent evaluators can reconcile differing views of the models’ relative strength as more detailed test results emerge.
Be the first to comment