Meta Debuts Muse Wonder, a Paid Frontier Model for Agents

Meta released two new models on Monday, Muse Wonder and Muse Wonder Full, calling them its most capable systems yet and pitching them straight at long-running agent work: insurance claims, lease reviews, personal finance, and coding. Muse Wonder is available now through the Meta Model API and Muse Code, while the larger Muse Wonder Full is limited to Muse Code subscribers. Meta says Wonder will soon power its consumer Muse agent, reaching paying subscribers first.

What Meta Announced

The company describes the pair as the second generation of its rebuilt AI stack, following Muse Spark from its Superintelligence Labs division. Two capabilities carry the pitch. Active Memory lets a model carry context across separate conversations and pull back whatever is relevant to the task at hand. Active Focus keeps an agent pushing toward a goal when a task gets hard: trying other approaches, recovering from errors, and checking in before doing anything irreversible.

Meta also claims the models trade blows with the current frontier. The Verge reports the company points to internal benchmarks against OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Fable 5.1, though Meta has not published an independent comparison with Google's Gemini 4 Argon.

The Paywall Is the Story

Access is the sharper change. CNBC reports that using Muse Wonder inside Meta's personal agent app requires a subscription at $20 or $100 per month. Meta AI chief Alexandr Wang said the model, developed under the codename Watermelon, was trained specifically to stop stalling: "When you do real-world tasks, you're going to get roadblocks... We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user."

Meta's own announcement says its strongest cyber-misuse safeguards ship only with the paid model for now, an unusual split that puts the safest configuration behind a billing wall.

Why It Matters for Builders

• Agent reliability is now the product. Memory across sessions and the stubbornness to retry a failed step separate a demo from something you can leave running, so test those two behaviours rather than raw leaderboard scores.
• Vendor lock-in lands on the runtime. Wonder Full is subscriber-only inside Muse Code while standard Wonder ships on the Meta Model API, so teams building on Meta should confirm which tier their workload needs before committing.
• Price the safety tier. If hardened guardrails sit behind the $20 and $100 plans, agent platforms will have to decide whether paying for the stricter configuration becomes a line item.

Meta's release also sharpens the subscription race: OpenAI, Anthropic, and Google are all pushing paid agent tiers over flat model access. For teams wiring agents into real workflows, the useful question is no longer which model tops a leaderboard, but which one keeps working after the first thing goes wrong, the theme running through Sunday's roundup.