Meta Launches Muse Wonder and Muse Wonder Full, Its Strongest AI Models Yet

Meta has released two new AI models, Muse Wonder and Muse Wonder Full. The launch came on Monday 12 October 2026 in the United States, which is 13 October in New Zealand. Meta calls them its most powerful models so far and says they are meant to compete directly with OpenAI and Anthropic. Muse Wonder is available now through Meta's Model API and its Muse Code tool. The larger Muse Wonder Full is limited to Muse Code subscribers.

Meta's own benchmarks claim the models top or match the latest offerings from OpenAI and Anthropic on personal agent, memory, and coding tests, according to Axios. Axios reports no specific scores, and independent testing is still pending. Meta also says the models perform strongly in law, finance, and health. One outside figure exists so far. VentureBeat, citing the Artificial Analysis Intelligence Index, puts Muse Wonder Full at 56. That ties Claude Sonnet 5.5 and sits below Claude Opus 5.5 at 58.

Two features stand out in Meta's description. Active Memory keeps information about the user across conversations. Active Focus keeps the model working on a difficult task when it hits an obstacle, and asks for confirmation before critical steps, according to ITdaily's summary of the launch. Both models accept text, images, video, and audio as input and output text. The context window is one million tokens.

ITdaily reports Muse Wonder API pricing at $4 per million input tokens and $20 per million output tokens. Axios does not mention pricing, so that figure rests on one source. It matches the standard rate Google will charge for Gemini 4 Argon from 1 December, which VentureBeat reports. The comparison is useful for developers choosing between the two, though neither company's rate has been tested against real workloads in this reporting.

Meta says the models were trained in Prometheus, its gigawatt-scale data center. Alexandr Wang, who leads Meta Superintelligence Labs, described them as "true frontier models, competitive with the best models in the world." He also said Meta is "still committed to open source." Meta has not said whether any Wonder model will have open weights. That silence stands out because the company recently released open weights for its earlier Muse Glimmer model.

On safety, Meta says Muse Wonder has its strongest protections yet against misuse for chemical, biological, radiological, and nuclear weapons. ITdaily reports that, without extra measures, Meta could not rule out Muse Wonder Full reaching the high-risk threshold for chemical and biological capabilities. After safety training and system-level classifiers, Meta rates all risks as moderate or lower. A model card was slated for release on launch day, with a fuller report to follow. Those ratings are Meta's own, and no independent group has checked them yet.

Meta's shares rose about 4% in afternoon trading after the announcement, according to Reuters. Mark Zuckerberg has committed billions of dollars to AI over the past year to close the gap with rivals, and new subscription tiers could help pay for that spending. Reuters also notes that privacy concerns remain around Meta's Muse agent.

Several questions remain open. Meta has not disclosed parameter counts. It has not said whether the models will be open weight. Independent benchmarks have not yet confirmed the company's claims. Pricing for Muse Wonder Full is not given beyond the subscription requirement, and the timing of the fuller safety report is unclear. It is also not yet known how quickly Muse Wonder will reach the broader Muse assistant for non-paying users.

For developers, the practical step is to run your own prompts through Muse Wonder and compare the outputs with the models you already use. Results that come back as JSON can be tidied with the free formatter at /en/json-formatter.