Meta puts its newest assistant model behind a paywall
X adds Cursor to its subscription, but third-party models carry a separate bill and AI usage is shared.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:004:27
Read transcript
Meta has put its newest assistant model behind a subscription. Muse Wonder, launched October 12, is available in the otherwise-free Muse app only to users on a $20 or $100 monthly plan. Meta calls it the company’s most advanced model for digital assistants, built to get further through tasks like booking dinner or checking availability—and to try another approach when it hits a roadblock instead of stopping to ask for help. That makes the launch a test not just of model capability, but of who gets access to it. Meta says Muse Wonder has its strongest safeguards yet against cyber misuse, but those protections are currently limited to the paid model. The company is therefore tying access to its stated strongest safeguards to a paid product, as it looks for AI revenue beyond advertising. Developer access has its own tiers. Muse Wonder is available through Meta’s paid API, though Meta has not disclosed the price. AI chief Alexandr Wang says it will cost more than standard subscriptions because the model is much larger, but be priced “meaningfully below” leading OpenAI and Anthropic models. The more powerful Muse Wonder Full is restricted to Muse Code plans costing $50 or $200 a month. The practical question is what users get for paying: a more capable agent, stronger safeguards, or both—and how reliably it can finish work without handing the task back. Watch for the API price and evidence of how the model performs beyond familiar errands. Across today’s stories, access, human oversight and operating cost matter as much as a model’s headline capability. That question of how much to trust an agent is playing out on the Formula One track, too. Oracle Red Bull Racing says its AI strategy tool has moved beyond looking up rules to recommending race-day actions. Team executive Jack Harington described a 2025 Italian Grand Prix example: Max Verstappen cut a corner while defending his position, and the agent advised giving the position back to avoid a potential penalty. The team followed that advice; Verstappen later overtook his rival. The tool was developed with Oracle. Harington says it weighs more than what is happening on track, including off-track developments and the championship picture. But he also stresses that engineers and strategists remain involved. The point, he says, is to help them make decisions, not replace them. As agents take on higher-stakes recommendations, the useful measure is not just whether they can produce an answer, but whether people can judge when to act on it. That gap between an agent’s answer and its value at work is central to Andrej Karpathy’s argument about how people experience AI. In an October 11 post, he says a brief chatbot exchange reveals little about systems used for demanding, long-running tasks—especially coding and mathematics, where results can be checked. Karpathy’s rough “napkin math” puts professional users of leading models in those fields at about 20 million, and people with access to internal frontier systems at roughly 5,000. He contrasts those groups with billions who have barely encountered large language models or use free chatbots casually. Those are estimates, not a measured census. He also describes internal systems collaborating over weeks on major software efforts, and says human review is beginning to fall behind. Those claims about internal capabilities are his account, not independently verified results. The useful takeaway is to ask what task, access level and verification standard a capability claim actually refers to—not to assume a casual chat is a meaningful test. And that need to test specific work, rather than rely on impressions, is also behind a new 3D benchmark from Artificial Analysis. AA-3D compares 21 models turning written briefs into explorable scenes. Each model can write code, render its work, inspect images and revise it, with a run capped at three hours or 250 agent turns. A panel compared scenes made from the same brief, producing preference-based Elo rankings. Claude Sonnet 5.5 had the highest launch score, but four models from Anthropic and OpenAI were statistically tied at the top. That ranking measures the panel’s preferences, not a universal definition of quality. Cost and speed tell a different story. Among the top five, GPT-6 Sol had the lowest mean cost, $3.46 per scene, and was quickest across the six tasks shown. So the shared lesson is to watch the fine print: who can access a system, where humans stay in the loop, and what the work costs to run. A leaderboard—or a subscription price—cannot answer all three by itself.
