Meta launches first frontier models, opens Wonder to developers via API

Summary

• Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, its first frontier-class models.
• Meta says Wonder matches Fable 5.1 and GPT-6 Astra on benchmark tests it ran itself.
• API access at $4 per million input tokens puts Meta directly against Anthropic and OpenAI.

The latest

Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, the company's first two models in the frontier tier, built around persistent memory and long-horizon task execution. Meta framed the launch as the next rung on its ladder toward what it calls personal superintelligence. Muse Wonder is live now through the Meta Model API; the larger Full variant is not.

Details

• Active Memory: According to Meta, the models retain important context about a user across separate conversations and pull back the details relevant to the task at hand. The company positions this as a core differentiator rather than a feature, and says Wonder beats Fable, Opus and Astra on the MemoryBench benchmark in its own testing.
• Active Focus: Meta says the models keep working on complex goals rather than stalling: trying alternative approaches when blocked, adapting to new information mid-task, and asking for clarification before taking consequential actions. That last behavior is the guardrail Meta highlights for agentic deployments.
• Domain claims: Meta cites expertise in law, finance and health, with examples such as parsing a lease agreement or interpreting financial and medical information. It also claims frontier-level coding ability spanning small apps through complex software, and better generalization to unfamiliar tasks, which it attributes to larger and stronger base models.
• The benchmarks: Meta names three specifically: Harvey's Legal Agent Benchmark (LAB), FocusBench and SWE-Marathon, and says the Wonder family posts leading results on key measures for business agents, memory and coding. These figures come from Meta's own testing and have not been checked by an independent party.
• The pricing: TechCrunch reported API usage at $4 per million input tokens and $20 per million output tokens, placing the models in the commercial tier where developers building agents choose between vendors on cost per task, not demo quality.
• Availability split: Muse Wonder is available now through the Meta Model API and through Muse Code. Muse Wonder Full is restricted to Muse Code subscribers and is not offered through the API. Meta says Wonder will soon power experiences in the Muse app, starting with subscribers, without naming a date.
• The safety claim: Meta describes these as its most aligned models to date, with its strongest controls yet against chemical, biological, radiological and nuclear (CBRN) misuse and cyber misuse, plus higher robustness against adversarial prompting. The full safety report is listed as coming soon.
• Where it was trained: Meta says the models were trained at its Prometheus data center in central Ohio, the facility the company has tied to its frontier-scale compute buildout.
• Rival comparison: Meta told TechCrunch that Wonder's performance is on par with Fable 5.1 and GPT-6 Astra on standardized tests, a direct positioning statement against the two labs that currently dominate commercial agent deployment.

Background

Meta had previously shipped open-weight and consumer-facing models but no frontier-tier system of its own. Muse Wonder is the first release carrying the Superintelligence Labs branding into that tier, alongside the Muse Code and Muse app product lines.

Between the lines

The pricing and API availability matter more than the benchmark table. By putting Wonder on a paid API at $4 and $20 per million tokens while keeping the larger Full variant inside Muse Code, Meta is selling the mid-tier to outside developers and reserving the top tier for its own subscription funnel. The benchmark results Meta cites are self-run, so the competitive claim against Fable and Astra rests on the company's own numbers.

What's next

Watch for the full safety report Meta says is coming, independent replication of the LAB, FocusBench, SWE-Marathon and MemoryBench results, and the start of Wonder-powered features in the Muse app for subscribers.

Source: Meta Superintelligence Labs, TechCrunch