Meta Launches Muse Wonder, Its First Frontier Model, Priced at $4 per Million Input Tokens

News Summary

Meta has released Muse Wonder and Muse Wonder Full, which TechCrunch describes as the company's first attempt at a true frontier model. The models come from Meta Superintelligence Labs. They were announced on Monday, October 12, 2026, and TechCrunch's report was timestamped 10:56 AM Pacific Time. Meta calls the release "the next step on our scaling ladder and a major leap toward personal superintelligence." Most of the details below come from TechCrunch's coverage and Meta's own materials. At the time of writing, we found no independent reporting that confirms the benchmark claims.

What Meta Released

There are two models: Muse Wonder and Muse Wonder Full. They are larger than Muse Spark, the lightweight, low-latency model Meta launched earlier. Meta says they are more capable, with higher benchmark scores than Spark and stronger results on complex tasks such as coding.

Muse Spark debuted on April 8, 2026, as the first model from Meta Superintelligence Labs. Secondary coverage at the time described it as a closed-weights, multimodal reasoning model. Meta acknowledged gaps in long-horizon agentic tasks and coding, which are the areas the Wonder models are meant to improve.

Active Memory as the Headline Feature

Meta's pitch centers on "active memory," the ability to retain and synthesize information about a user over long periods. Kavya Kopparapu, a research scientist at Meta, said in a launch video: "We put a tremendous amount of effort into making sure your model can remember what's important to you."

For learners and developers, this is a notable design goal. Most assistants treat each conversation as largely separate. A model that builds and updates a persistent picture of a user's preferences and projects has to decide what to keep, what to summarize and what to discard. That is a research problem in itself.

Benchmark Claims

According to the model card, Meta's own testing shows Wonder and Wonder Full scoring comparable to Fable 5.1 and GPT-6 Astra. On MemoryBench (arXiv 2510.17281), a benchmark for continual learning and memory construction, Muse Wonder reportedly scores higher than Fable, Opus or Astra.

These figures are self-reported. Vendor-run benchmarks often differ from independent evaluations, and memory benchmarks are still young. Independent testing from third-party evaluators will be the real check.

Availability and Pricing

Access is split between the two models:

• Muse Wonder is available through the Meta Model API.
• Muse Wonder Full is limited to Muse Code subscribers and is not served through the API.
• Both models will power the Muse agent, which Meta introduced in September 2026.

API pricing for Muse Wonder is $4 per million input tokens and $20 per million output tokens. TechCrunch notes this is roughly in line with GPT-6 Sol or Opus 5.5.

Training Infrastructure

The models were trained at Meta's Prometheus data center in central Ohio, which came online earlier in 2026. TechCrunch does not report parameter counts, architecture, context window or training data details, and Meta had not published them in the material we reviewed.

Why It Matters

Alexandr Wang, founder of Scale AI, restructured Meta's AI efforts in August 2025. Since then Meta has shipped Muse Spark and the Muse agent, but not a model aimed at the top of the field. Wonder is the first release positioned that way.

Two things are worth watching. The first is whether independent evaluations confirm the benchmark parity Meta claims. The second is whether active memory works well in daily use, since that is the feature Meta is betting on.