Meta unveils Muse Wonder models, claiming a return to the AI frontier

Meta on October 12, 2026 released Muse Wonder and Muse Wonder Full, two agent-focused AI models with a 1 million token context window. They are the most ambitious output so far from Meta Superintelligence Labs, the unit Mark Zuckerberg set up last year. The company described the pair as "the next step on our scaling ladder and a major leap toward personal superintelligence."

Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens on the Meta Model API. The more capable Muse Wonder Full is offered through plans for Muse Code, Meta's AI-assisted programming tool. Meta said the top version will also power features across its platforms, including its Muse personal agent.

How it stacks up

According to Meta, the largest version beats Anthropic's Claude Fable 5.1 on the Artificial Analysis Intelligence Index, a popular aggregate measure of model ability, and sits just behind Claude Opus 5.5. Meta also claims it tops OpenAI's GPT-6 Astra and both Claude models on certain specialised tests, including legal document work and software engineering. Wonder Full is reported to lead the DeepSWE and SWE-Marathon benchmarks. Benchmark scores, however, do not always translate into real-world performance.

The models introduce two features called Active Memory and Active Focus, intended to keep an agent anchored to its assigned goal and stop it from losing track during long jobs such as deep research or large coding projects, where drift can cause costly, compounding mistakes. Meta did not reveal how the models were built or trained.

Safety and the road back

Meta says it invested heavily in training Muse Wonder to refuse dangerous requests, such as bioweapon instructions or hacking assistance, and that it scores well on tests of resistance to such prompts. It did not say whether Muse Wonder-powered agents misbehaved in testing, a live concern after recent incidents involving rogue agents from other labs. Zuckerberg wrote on X in September 2026 that the Muse launch had been delayed for safety work, and Meta published a framework for scaling capabilities safely in early October.

Meta once led open-weight AI with its Llama series, but slipped behind Anthropic, OpenAI and Google after several underwhelming releases. Zuckerberg responded by pouring billions into the new lab and recruiting former Scale AI chief executive Alexandr Wang to run it, an effort marked by some researcher departures. The lab's first model, Muse Spark, arrived in April 2026, followed by the small open-weight Muse Glimmer. Muse Wonder suggests Meta is competitive again in an intensifying race.