Meta released Muse Wonder and Muse Wonder Full on October 12, describing them as its most capable models to date and its first frontier-class releases. The company said the models deliver state-of-the-art performance on personal and professional agentic tasks, including Active Memory and Active Focus features that carry context across conversations and recover from errors.

Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is limited to Muse Code subscribers. Meta said the models will soon power its Muse personal agent, starting with subscribers. Artificial Analysis scored Wonder Full at 56 on its Intelligence Index v4.3, tying Anthropic’s Claude Sonnet 5.5 and surpassing OpenAI’s GPT-6 Astra at 53.

The release follows a 14-month rebuild of Meta’s AI organization under Alexandr Wang. Mark Zuckerberg posted the announcement on X at 12:55 pm Eastern. Meta’s shares rose about 4% in afternoon trading on the day of the launch. A full safety report is still pending.

Key takeaways

What Meta’s Muse Wonder pricing and access actually mean

The API pricing sits near GPT-6 Sol and Opus 5.5 levels. Cached input drops to $0.20 per million tokens. Wonder Full stays inside Muse Code for now, so most external developers will see only the standard Wonder model. Meta trained the pair at its Prometheus data center with a 1 million-token context window.

Independent benchmarks from Artificial Analysis show the largest gains on agentic evaluations such as Terminal-Bench 4.0. Meta’s own model card places the scores in line with Fable 5.1 and GPT-6 Astra. The company has not yet published the full safety evaluation set referenced in its Superintelligence Scaling Framework.

What to watch next

Rollout timing for Muse Wonder inside the Muse agent will show how quickly Meta converts the model into a subscription product. Developer uptake through the Model API will test whether the pricing and scores pull traffic from OpenAI and Anthropic endpoints. The pending safety report will clarify how Meta handled the same categories of unintended actions that prompted Anthropic to cut live internet access from its internal evaluations earlier in the month.

For creators building content systems, the Active Memory feature is the practical change. It reduces the need to restate context across long sessions, which is the same constraint that has shaped most AI content stacks in 2026. See How AI is Reshaping the Content Creator Stack in 2026 for the broader workflow picture.