Meta claims new Muse Wonder AI models rival Axiom, Paradox

Muse Wonder and Muse Wonder Full are Meta’s newest artificial intelligence models. Meta describes them as its most capable models so far. The company says they point toward “personal superintelligence,” meaning highly capable AI that can assist an individual across conversations and demanding tasks. The phrase is Meta’s vision, not a technical measurement defined in the article.

Their main advertised features are Active Memory and Active Focus. Active Memory carries relevant information from one conversation into another. Active Focus helps an agent keep working on difficult tasks instead of losing track. Together, these features aim to make the models more useful for extended, multi-step work.

The models are already being positioned against leading systems. Artificial Analysis scored Wonder Full at 56 and standard Wonder at 52. Wonder Full is available only through Muse Code’s $50 or $200 monthly power-usage tiers, making advanced access costly.

Artificial Analysis gave Meta’s models different scores on its Intelligence Index. At maximum reasoning effort, Muse Wonder Full scored 56. The standard Muse Wonder scored 52. These numbers place the two versions close together, with Full four points ahead of standard Wonder.

Wonder Full scored above both Paradox Fable 5.1 and Axiom Astra, which each received 53. It therefore led those two models by three points. Standard Wonder’s score of 52 was one point below them. Paradox ranked highest among the models listed, with a score of 58.

The figures suggest strong performance, but they are not universal proof of superiority. Bindu Reddy of Abacus.AI disputed Meta’s claims and said Muse Wonder was “nowhere near Fable or Astra level.” Reddy said detailed benchmark results would follow, showing why methodology matters.

Active Memory and Active Focus are Meta’s names for two new model capabilities. Active Memory carries relevant information across conversations. That means the model can use earlier context later instead of treating every conversation as completely separate. Active Focus helps an agent sustain work on difficult tasks, keeping attention on a demanding objective for longer.

For example, a user could discuss a project in one conversation and return to it later. Active Memory could preserve relevant details. If the project requires several steps, Active Focus could help the agent continue working through the task rather than stopping after a quick response. The article does not describe the underlying software or model design.

These capabilities matter because they target continuity and persistence, not just one-off answers. Meta presents them as part of a move toward “personal superintelligence.” Their real-world reliability, limits, and performance compared with rival systems remain disputed, since independent evaluator Bindu Reddy challenged Meta’s claims.

If Meta’s performance claims are accurate, developers could gain a credible alternative to leading Axiom and Paradox models. That could encourage more competition, give teams another system to test, and reduce dependence on any single provider. Users might receive assistants that remember useful context and persist through harder tasks.

The practical effect would depend on price and access. Standard Wonder charges $4 per million input tokens, while Wonder Full is limited to Muse Code subscribers paying $50 or $200 monthly. Developers could choose the cheaper model for some workloads and reserve Full for tasks where its higher Intelligence Index score matters.

The outcome is not settled. Artificial Analysis ranked Wonder Full above Axiom Astra and Paradox Fable 5.1, but Bindu Reddy disputed that conclusion. If later evaluations confirm Meta’s claims, adoption could grow. If they do not, developers may continue favoring rival models despite Meta’s headline scores.

Independent evaluators can disagree because “capability” is not one single property. A test may emphasize coding, reasoning, factual accuracy, speed, or instruction-following. Different question sets and scoring rules can reward different strengths. The article does not provide the full methods behind either evaluation, so the numbers cannot be compared without that context.

Reasoning settings can also affect results. Artificial Analysis scored Wonder and Wonder Full at maximum reasoning effort. That setting may give a model more time or computation than another evaluation allows. An evaluator using different limits, prompts, or tools could produce a different ranking. Model versions and test conditions can matter too.

The disagreement appears directly in the article. Artificial Analysis gave Wonder Full 56, above Fable and Astra at 53. Bindu Reddy said Muse Wonder was “nowhere near” those models and promised detailed benchmark results. More transparent, repeatable testing would help explain the gap.

Tokens are small pieces of text that AI models process. Providers commonly charge by the number of tokens sent into a model, called input tokens, and sometimes by generated output tokens as well. One million tokens is a billing unit. In this article, Meta charges $4 per million input tokens for standard Wonder.

A higher price can still make sense when a model delivers better results. A customer may value stronger reasoning, fewer mistakes, longer task persistence, or less human correction. Those benefits can outweigh the model’s fee for important work. Price also depends on access: Wonder Full is available through $50 and $200 monthly Muse Code tiers, not simply the listed standard rate.

The article shows a competitive price spread. Standard Wonder matches Paradox at $4 per million input tokens. Axiom Astra costs $10, while Paradox Fable 5.1 costs $50. Customers may compare price, capability, reliability, and access together.

Large language models turn text into tokens, such as word pieces or punctuation. During training, they process huge collections of examples and adjust internal parameters to learn language patterns, relationships, and problem-solving behaviors. When answering, the model predicts likely next tokens repeatedly until it forms a response. The article does not describe Muse’s training data or architecture.

Computing power supports both training and answering. More inference computation can allow a model to consider a problem for longer or explore more reasoning steps. That helps explain why Artificial Analysis reported scores at maximum reasoning effort. Active Focus is described as helping an agent sustain difficult work, but the article does not say exactly how it achieves that.

Active Memory addresses continuity rather than basic token prediction. It carries relevant information across conversations, giving later responses useful context. Together, computation, learned patterns, and retained context can support longer tasks. Their effectiveness still needs independent testing, since evaluators disagree about Meta’s rankings.