Meta’s Muse Wonder Puts Zuckerberg Back in the Frontier AI Race

Key Notes

• Muse Wonder Full scores 56 on Artificial Analysis’s Intelligence Index, placing Meta in the group behind Opus 5.5’s leading 58 rather than in an uncontested second place.
• Meta’s models show specific strengths in legal-agent work and sustained coding, while OpenAI, Anthropic, Google and Grok compete through different combinations of capability, price and product access.
• Wonder is available through the API at $4 per million input tokens and $20 per million output tokens, while higher-scoring Wonder Full requires an eligible Muse Code subscription.

Mark Zuckerberg now has a stronger answer to the question that has followed Meta’s AI ambitions: can the company build a model that belongs alongside the best from OpenAI, Anthropic and Google? With Muse Wonder and Muse Wonder Full, the answer is increasingly credible. Meta has put a new contender near the top of an independent intelligence leaderboard and into the market for professional AI agents.

The October 12 launch makes the competition more crowded. OpenAI has Astra and its broader GPT family, Anthropic has Claude, Google is advancing Gemini, and Elon Musk’s Grok is competing on coding and cost. Zuckerberg’s Muse now gives that group another serious rival, particularly for tasks that require an agent to remember context and work through a long project.

Our launch coverage examined the models’ features and access. The larger significance is what happens when another well-funded company becomes a credible option for demanding work. Meta does not need to win every benchmark to change purchasing decisions; it needs enough capability, a useful product and a reason for people to switch.

A 56-Point Score Puts Meta Near the Front

On the Artificial Analysis leaderboard checked October 13, Muse Wonder Full at max effort scores 56 on the Intelligence Index. Claude Opus 5.5 at max with fallback leads at 58. Wonder Full shares its displayed score with Sonnet 5.5 at max with fallback and Opus 5.5 at xhigh with fallback.

That places Meta in the group immediately behind the highest score, rather than establishing an uncontested second-place finish. The ordinary Muse Wonder model scores 52, below Full. The distinction matters because the two versions have different access arrangements.

The selected configurations below show the competitive picture on that index at publication. They compare particular models and settings, rather than ranking companies or measuring every possible use case.

A broad index is useful for identifying models worth testing. It cannot settle whether a particular assistant will complete a company’s migration, handle its documents or follow its approval rules. The commercially meaningful change is that Meta has earned a place in that evaluation, rather than asking developers to choose Muse on its future potential alone.

Zuckerberg’s Bet Combines Compute With Distribution

The launch is a milestone for the rebuilding effort around Meta Superintelligence Labs. Meta says the Wonder models were trained at its gigawatt-scale Prometheus data center. That connects the release to a physical infrastructure strategy, alongside the research work needed to turn computing capacity into useful models.

The financial scale is substantial. In its July earnings release, Meta reported $31.08 billion in second-quarter capital expenditures, including finance-lease principal payments, and forecast $130 billion to $145 billion for 2026. Those are company-wide figures, not a disclosed training bill for Wonder or a separate budget for the superintelligence lab.

Meta also reported an average of 3.60 billion daily active people across its family of apps in June. That gives it an enormous potential distribution channel, although those users should not be counted as Muse customers. Turning access to an audience into habitual use of a professional assistant is a separate challenge.

Zuckerberg’s stated vision emphasizes personal superintelligence that people direct toward their own goals. Wonder gives that ambition a more concrete product proposition. A model that remembers useful context and continues working could become a recurring part of someone’s day, rather than a chatbot they consult only when they have a question.

OpenAI and Anthropic Set Different Hurdles

OpenAI’s challenge to Meta extends beyond any single leaderboard position. Its Astra documentation positions GPT-6 Astra for complex reasoning, coding, computer use, research and document creation. The model supports a 1,050,000-token context window and adjustable reasoning effort, with standard API rates of $10 per million input tokens and $50 per million output tokens.

For Meta, competing with that range means making Muse useful across a workflow, not just producing a stronger answer on one test. Users may need to investigate a problem, inspect documents, write code and deliver a finished result within the same project. Our coverage of OpenAI Dots shows how the competition is moving toward persistent assistants around the underlying models.

Anthropic remains a demanding comparison for both capability and economics. Its Opus 5.5 announcement lists $4 per million input tokens and $20 per million output tokens, matching Wonder’s published base rates. Anthropic emphasizes complex coding, knowledge work and improved efficiency, while its max configuration still leads the independent index shown above.

This means Meta cannot rely on a simple claim that it delivers frontier AI at a price Claude cannot match. The stronger argument has to concern the work Muse does particularly well: remembering relevant information, sustaining an engineering task or producing a useful professional deliverable. For customers already comfortable with Claude, those advantages need to be visible in their own projects.

Google and Grok Keep the Race From Becoming a Duel

Google’s Gemini 4 Argon announcement describes a phased release, with introductory pricing of $2 per million input tokens and $10 per million output tokens. Google says it is expanding access while gathering feedback and iterating on safeguards. A strong evaluation result and broad availability are therefore different stages of the release.

That creates another competitive constraint for Meta: a rival can put downward pressure on pricing even while its newest flagship is still reaching users. We explored the same distinction in our Argon analysis. Buyers need to distinguish the model being advertised from the version they can actually use in a supported workflow.

Grok is pursuing its own route. The official Grok 4.7 release emphasizes longer coding and knowledge-work tasks, improved verification and availability through Cursor, Grok Build and the API. Its published starting rates are $2 per million input tokens and $6 per million output tokens; a faster variant costs more.

Those prices do not make Grok interchangeable with Wonder, Claude or Astra, but they make cost part of the contest. A team may choose a lower-priced model for frequent assignments and reserve another for difficult work. As our Grok coverage noted, coding performance and efficiency are central to that pitch. Meta is entering a market where customers can compare several combinations of quality, speed and expense.

Where Muse Makes Its Strongest Case

Meta’s model card points to strengths in particular professional tasks. It reports a Harvey LAB-AA legal-agent score of 13.1 for Wonder Full, compared with 8.6 for Astra and 4.2 for Opus 5.5. On SWE-Marathon, it lists Full at 57.5 using xhigh effort, against 42.5 for Astra and 46.3 for Opus.

The same report shows limits to Meta’s lead: Full trails both rivals on Terminal-Bench 4.0-AA, and Opus scores higher on ProgramBench. The report combines internal evaluations, third-party results and comparisons involving different agent systems. Its strongest evidence supports a competitive model with identifiable strengths, rather than a universal winner.

Meta’s product emphasis is equally important. Active Memory is designed to retrieve relevant context across conversations, while Active Focus supports continued work, alternative approaches and clarification when needed. For an assistant handling a long assignment, remembering a constraint at the right moment can matter as much as solving an isolated reasoning problem.

Consider a software project with a fixed budget, an existing codebase and a requirement to check before changing a production system. Useful persistence means keeping those constraints while navigating errors and revisions. The attraction is less repeated explanation and supervision; the test is whether the agent preserves the right boundaries while completing the work.

The Best Score and the Available Product Are Different

Meta’s developer page lists Wonder through the Meta Model API and Muse Code, with a one-million-token context window. Its rates are $4 per million input tokens, $0.20 for cached input and $20 per million output tokens. Wonder Full is subscription-only through Muse Code’s $50 Power Usage and $200 Maximum Usage plans.

The distinction prevents a misleading comparison: the 56-point Full result is not the score of the model offered at Wonder’s API prices. Developers integrating a model into their own service and subscribers using Meta’s coding application are buying different access paths. Full’s subscription price also does not translate directly into a per-token rate or unlimited usage.

Published token rates are only the starting point for cost comparisons. Astra, for example, applies higher rates to prompts beyond 272,000 input tokens, while processing modes and tool charges can also affect bills. More generally, an agent’s total cost depends on how much work it performs, how often it retries and how much review its output requires.

A cheaper request can become expensive if it produces work that needs substantial correction. A higher-priced model can justify its cost when it completes a difficult assignment reliably. This is why the next round of competition will depend on completed tasks and the time people save, alongside the intelligence scores that attract attention at launch.

A More Crowded Frontier Is Good News for Buyers

Wonder’s arrival gives developers and businesses another supplier to evaluate for demanding work. It can strengthen their negotiating position and reduce the need to build every workflow around one company’s models. That does not eliminate switching costs: prompts, tools, evaluation suites and operational procedures still need to work with a new provider.

The larger opportunity for Meta is to connect competitive intelligence with a product people trust enough to delegate to. Memory and persistence become valuable when users can inspect results, correct assumptions and control consequential actions. The same features become liabilities if an assistant carries an old mistake into a new task or keeps working beyond its authority.

Zuckerberg now has a credible frontier contender. OpenAI, Anthropic, Google and Grok still present different challenges, and none of those contests is settled by one release. Muse Wonder makes the race more consequential because Meta can now compete for serious work on current capabilities. Its next achievement must be turning that capability into users who keep choosing it after the launch excitement fades.

Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.