Meta Releases Muse Wonder AI Models, Pressuring Leading Rivals

Meta has released two new AI models—Muse Wonder and Muse Wonder Full—positioning them as its most capable systems yet and describing them as a step toward “personal superintelligence.” The company says the models introduce two new features it calls Active Memory and Active Focus, designed to help agents carry relevant context across conversations and stay on track when working through difficult problems.

The announcement arrives alongside independent benchmark measurements from Artificial Analysis, which reported that Wonder Full scored higher than the standard Wonder on its Intelligence Index—while also placing both models behind top performers in the same evaluation set. At the same time, Meta’s performance claims drew pushback from at least one independent evaluator.

Key takeaways

• Meta describes Active Memory as the ability to carry relevant information across conversations, and Active Focus as a mechanism to sustain progress on challenging tasks.
• Artificial Analysis scored Wonder Full at 56 and Wonder at 52 (both at maximum reasoning effort), while Claude Opus 5.5 scored 58 in the same index.
• Artificial Analysis data places Wonder Full above Claude Fable 5.1 and GPT-6 Astra (both 53), but still below Opus 5.5.
• Meta’s pricing for Wonder is $4 per million input tokens; Wonder Full is restricted to paid Muse Code “power” and “maximum” usage tiers.
• Abacus.AI CEO Bindu Reddy challenged Meta’s benchmark framing, saying Muse Wonder was “nowhere near” competing models and promising more detailed results.

Meta’s “Active Memory” and “Active Focus”

In a post on Meta Research, the company said the Muse Wonder family adds two capabilities intended to improve how AI agents function over time.

According to Meta, Active Memory is meant to keep and reuse information that remains relevant across separate conversations, rather than treating each exchange as entirely isolated. The goal is to make interactions feel more consistent and context-aware—particularly for tasks where prior decisions or user preferences matter.

Meta also described Active Focus as a way for the agent to maintain momentum on difficult work. Rather than simply answering in a single pass, Meta frames the feature as supporting sustained progress when problems are complex, potentially reducing the risk of losing direction mid-task.

How independent benchmarks place Wonder Full and Wonder

Independent evaluation platform Artificial Analysis published Intelligence Index scores for both models. It reported a score of 56 for Wonder Full and 52 for the standard Wonder, with both evaluated at maximum reasoning effort.

Artificial Analysis’ index also showed Wonder’s relative standing among other models it tested. The platform placed both Meta models above Claude Fable 5.1 and GPT-6 Astra, which it said scored 53 each. However, it found Claude Opus 5.5 ahead, with a score of 58.

For investors and model-adoption watchers, the practical implication is that Meta’s new releases appear to be incremental-to-competitive rather than clearly dominant—at least based on the single benchmark set Artificial Analysis chose to publish. Readers should also note that benchmark outcomes can vary depending on evaluation design, prompt format, and how reasoning effort is configured.

Dispute emerges over how to interpret Meta’s claims

Meta’s positioning did not go unchallenged. Bindu Reddy, CEO of Abacus.AI and a figure who runs his own model evaluation work, criticized the performance narrative on X. He wrote that Muse Wonder was “nowhere near Fable or Astra level” and said more detailed benchmark results would follow, pointing to a disagreement over either relative performance or the way benchmarks were presented.

While Reddy did not provide counter-scores in the post excerpted in the coverage, the pushback highlights a recurring issue in the AI landscape: model announcements often come with strong claims, but independent scrutiny can produce materially different interpretations depending on evaluation methods. The next test for readers is whether subsequent third-party evaluations converge with Artificial Analysis’ reported numbers or support Reddy’s concerns.

Pricing and access: Wonder is more broadly priced, Full is tier-gated

Meta also disclosed how the models will be priced and sold. According to the reporting, Muse Wonder charges $4 per million input tokens. That figure is described as matching Claude Opus 5.5’s input-token pricing while undercutting GPT-6 Astra and Claude Fable 5.1, which are reported at $10 and $50 respectively.

Access to Wonder Full is more restrictive. It is available only to subscribers of Muse Code’s usage tiers, described as $50 per month for “power usage” and $200 per month for “maximum usage.” This tiering matters for developers evaluating whether they can realistically test the “Full” version under normal budget constraints, especially for teams building agentic workflows that may require repeated calls.

For users comparing model value beyond raw benchmark scores, these commercialization details are often as important as the performance numbers. A model that scores higher but costs more—or is constrained to specific tiers—can still be less practical for many applications.

What to watch next

Beyond the initial benchmarks, the key question will be whether Active Memory and Active Focus translate into more reliable agent behavior in real workloads, and whether additional independent evaluations align with Artificial Analysis’ Intelligence Index scores or with Abacus.AI’s critique. As more benchmark results and practical reports emerge, the market will get a clearer view of where Muse Wonder and Muse Wonder Full truly sit relative to the strongest competing systems.