Meta finally has an answer to the question that's dogged it all year, and it's not a benchmark score — it's a price tag.

The company announced Muse Wonder and Muse Wonder Full on Monday, its "most capable models to date," built for agentic work: models judged on whether they can finish a task, not whether they can chat. Muse Wonder is available now through the Meta Model API and Muse Code; the heavier Muse Wonder Full is reserved for Muse Code subscribers, with a phased rollout of Muse features to paying subscribers of the Muse app.

The tell is that Meta's AI chief Alexandr Wang led with pricing, not capability. He said the models would cost "meaningfully below" what OpenAI and Anthropic charge — a direct, named attack on the two labs everyone else is too cautious to call out in a launch post. Reported plan tiers run $20 to $100 a month for the Muse app and $50 to $200 for Muse Code, though Meta hasn't published per-token API pricing yet, so the real developer cost comparison has to wait.

The substance is in the two named features: Active Memory, designed to retain context across conversations spanning days or weeks, and Active Focus, which keeps models grinding through multi-step tasks — trying alternate approaches, asking for clarification instead of guessing. The demo cases are deliberately unglamorous: filing an insurance claim, interpreting a lease, building an app. Wang's implicit argument is that agents are judged on whether they can book the dinner reservation without failing, and that's a harder trust problem than any reasoning benchmark.

Skepticism is warranted. TechCrunch framed the launch as Meta trying to prove it has merely caught up to the frontier rather than leapt past it, and Engadget's take was nearly identical: Meta has closed the gap, not overtaken anyone. The internal codename was reportedly "Watermelon" — fitting, since this is a red-fruit moment: big, splashy, and still waiting to prove there's juice inside.

But dismissing this as catch-up misses the strategy. The agentic AI fight of 2026 isn't about who has the smartest model — it's about who can deliver a reliable enough agent at a price point that makes deployment economics work. Meta's real weapon was never the model; it's the distribution. OpenAI and Anthropic can't funnel users through Instagram, WhatsApp, and Facebook. If Muse Wonder is even close to competitive and meaningfully cheaper, price becomes the feature, and the frontier labs suddenly have to answer a question they've never faced: what happens when the best model doesn't win?

---

— via Tech Insider.