Meta enters the frontier: what Muse Wonder means for your work
Meta's first frontier AI models are here, priced like the leaders. What changed, what it costs, and when it actually matters for your business.
Meta has joined the frontier AI race. On October 12, its Superintelligence Lab released Muse Wonder and Muse Wonder Full - the company's most capable models to date, built to compete directly with OpenAI and Anthropic instead of sitting in the cheaper lane beside them. Here's what the launch really tells us, beyond the benchmark charts. What Meta actually released: Two models, not one. Muse Wonder is available through the Meta Model API. Muse Wonder Full - the stronger version behind the headline benchmark scores - is reserved for Muse Code subscribers and has no API access at all. Both power Meta's Muse agent. This is the first frontier-class release from the Superintelligence Lab, about fourteen months after Mark Zuckerberg reorganized Meta's AI teams under Alexandr Wang. The weights are closed - a clear break from the open Llama era. The benchmarks: parity, not supremacy: Meta's own charts put Wonder Full at 57.5 on the SWE-Marathon coding benchmark, ahead of Claude Opus 5.5 at 46.3 and GPT-6 Astra at 42.5. But read the footnote: Meta admits that wasn't a controlled comparison, and its model ran at maximum effort. Independent testing from Artificial Analysis tells a calmer story - Wonder Full scores 56 on the Intelligence Index, just behind Opus 5.5 at 58 and roughly level with Fable 5.1 and GPT-6 Astra at 53. The honest summary: Meta has caught up to the pack, not passed it. The two things actually worth noticing: First, memory. Meta's big pitch is "Active Memory" - the model retaining and building on what it knows about you across days and weeks, not just one chat. On Meta's MemoryBench test it outscores the rivals, though that's Meta's own benchmark. Paired with "Active Focus" for agentic work, this is Meta betting that the next advantage isn't raw intelligence but continuity. Second, the price tag. At $4 per million input tokens and $20 per million output tokens, Wonder costs exactly what Claude Opus 5.5 costs. The era of Meta as the cheap, open alternative is over - it's now charging frontier prices like everyone else. What this means for your work: For most businesses, nothing changes this week. But the direction matters. If your work involves long-running tasks where context carries over - ongoing projects, customer relationships, multi-step workflows - models with serious memory are the ones to watch, and Wonder is now a credible option alongside the incumbents. If you're already building on an API, Wonder gives you a fourth frontier supplier to compare, which is good for pricing leverage alone. Don't switch on benchmark scores; run your own tasks through two or three models and measure what actually matters to you - accuracy on your data, cost per task, and how often you need to correct it. The bigger picture: Fourteen months ago Meta wasn't in this conversation. Now OpenAI, Anthropic, Google, and Meta are in a four-way contest, with GPT-6 Astra landing September 3 and Opus 5.5 on September 22 setting the pace Meta is chasing. For anyone buying or building with AI, that competition is the real news - more capable models, more choice, and pricing pressure that benefits the customer. The catch-up phase is ending; the differentiation phase - memory, agents, continuity - is beginning.