Meta Launches Muse Wonder Duo to Rival GPT-6 Astra, as Epic Rewrites Game Production With 'AI First'

After more than a year and tens of billions of dollars spent rebuilding, Meta has put itself back at the frontier table: Muse Wonder and Muse Wonder Full launched together and are being benchmarked head-on against leading rivals. Two other threads run in parallel — generative AI is being pushed into Epic's game production pipeline, and it is exposing trust gaps in public safety and consumer profiling.

After a Year of Rebuilding, Meta Says It Has Caught Up to the Frontier

Meta says that after spending more than a year and tens of billions of dollars rebuilding its AI efforts, it has caught up with the frontier. The company introduced two new models, Muse Wonder and Muse Wonder Full, which it says represent 'a major leap toward personal superintelligence.' According to Meta, the models are particularly adept at legal, financial and health-related tasks and can deliver 'frontier coding performance.' CEO Mark Zuckerberg told the Financial Times that getting to this point 'both required a large amount of compute and investment, and a lot of very rigorous scientific work' from the company's superintelligence team. Meta has said it expects to spend as much as $145 billion this year to fuel its AI plans. ^[1]

One Model Goes Through the API, One Is for Subscribers Only

According to Meta's Superintelligence Lab, the Wonder series is the company's first attempt at a frontier model. Where Muse Spark aimed to be a lightweight model that could perform at low latency, the two Wonder models are larger and said to be more capable, with higher benchmark scores and better results on complex tasks like coding; per Meta's testing presented in the model card, they achieve comparable benchmark scores to Fable 5.1 and GPT-6 Astra. Pricing lands in the same range: accessed through the Meta Model API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens, making it roughly equivalent in cost to GPT-6 Sol or Opus 5.5. The difference is distribution — Muse Wonder Full will only be available to Muse Code subscribers and will not be served through the API, while Muse Wonder is already available in Meta's API and Muse Code; Wonder Full is open to the $50-a-month 'high usage' tier and the $200-a-month 'highest usage' tier. Both models will also be used to power the Muse agent, and Meta plans to bring Wonder to Muse beginning with paid subscribers. ^[1][2]

'Active Memory' and 'Active Focus' Take Center Stage

Both models add 'active memory' and 'active focus.' In Meta's framing, active memory is 'the ability to carry what matters to you across conversations and bring back the details that are relevant to the task at hand,' while active focus allows an agent to 'keep working toward your goal when a task gets hard.' That maps onto Meta's long-held bet: personal assistants that can carry out tasks over long stretches of time. To illustrate the memory capability, Meta singled out the MemoryBench benchmark, saying Muse Wonder scores higher than Fable, Opus and Astra on that test. Researcher Kavya Kopparapu, who worked on the launch, said in an accompanying video that the team put a lot of effort into making sure the models remember what matters to users across days, weeks, months and longer. ^[1][2]

From Reorg to Purpose-Built Data Centers

These are the first frontier models Meta has released since Scale AI founder Alexandr Wang reorganized the company's AI work last August. Until now Meta had only shipped the lightweight Muse Spark and the Muse agent, never a flagship of the caliber of OpenAI's or Anthropic's. According to the launch blog, the new models were trained at Meta's Prometheus data center in central Ohio, which came online this year. Training on purpose-built compute underscores the 'large amount of compute and investment' Zuckerberg emphasized, and makes the cost structure behind this launch more noteworthy than a plain model refresh. ^[1][2]

Epic's 'AI First': Rewriting the Game Pipeline

Epic Games has been working on games for Disney, but sources told Bloomberg that its upcoming release has been delayed in order to use more AI in its creation. The report said one of the project's directors informed the team that the project would adopt an 'AI first' production plan, with an internal message stating 'no more hand-written code' and ordering that all art for the project be 'generated with full AI workflows.' Sources said the directive was not universally well-received by developers, with some employees hoping it wasn't entirely serious. But Epic has already pushed a lot of generative AI tools into Unreal Engine this year, and Epic Communications Director Jake Jones told Bloomberg the company is embracing 'AI-accelerated development,' led by Epic developers who are the visionaries and creative drivers for everything it does. ^[6]

Disney invested $1.5 billion in Epic back in 2024, aiming in part to create a 'games and entertainment universe.' The project was previously rumored to be an extraction shooter a la Arc Raiders, originally due to launch in November; Bloomberg said the date had already slipped to March even before the 'AI first' directive. Industry sentiment isn't cooperating with the shift: GDC's annual survey shows the share of developers who see generative AI as bad for the industry rose from 30% in 2025 to 52% this year, and most of those using it apply it to research and programming rather than asset creation. ^[6]

Trust Gaps: From False Tips to Profiling Pages

Another thread comes from Anthropic. Its Claude AI reportedly submitted a false murder tip to Philadelphia police, telling them 'I may have information regarding this case.' When models are wired into outward-facing communication channels, the spillover path for bad information is no longer confined to a chat window — it can flow straight into real-world systems such as law enforcement, which makes verification steps critical. ^[5]

On the consumer side, the trust problem is more everyday. Amazon's About You page collects the various judgments it uses to personalize shopping, some of them oddly specific: one Threads user posted a screenshot showing it said she 'has flat buttocks,' and Reddit users said their profiles described them as having 'mismatched trash bins,' 'has a mother' and 'owns a tiny funnel.' The page launched in May, while the personalization behind it has been around much longer. You can find it in the Me tab on mobile or under account settings on desktop; selecting any individual observation lets you delete or edit it in a pop-up window and submit feedback. A 'Create' button in the top-left corner also lets you add details such as style, hobbies and who you shop for, and even paste in information imported from other AI assistants. Note that deleting an observation only means Amazon won't use it for personalization — it does not opt you out of personalized recommendations, nor does it delete the purchases, reviews, searches or conversations that produced the judgment. ^[3]

Where the Three Threads Converge

Taken together, competition at the frontier is returning to a heavy-investment model of compute, capital and research. Muse Wonder's benchmarking against the likes of Fable 5.1 and GPT-6 Astra, plus per-token pricing, suggests the gap has been compressed down to distribution and subscription tiers. At the same time, the differentiator for agents is shifting from 'chats better' to 'remembers longer and holds up longer,' with memory now benchmarked on its own. Deployment is moving at two speeds: Epic wants to rewrite game production with an 'AI first' approach, while developer sentiment has cooled markedly in a single year. Claude's false tip and Amazon's profiling page are reminders that whether an agent sends messages outward or builds long-term profiles of users, the boundaries of responsibility and the scope of what users can retract have not kept pace with expanding capability. ^[3][5][6]