Meta's Superintelligence Labs debuts Muse Wonder, its first frontier-class models
Meta Superintelligence Labs has released Muse Wonder and Muse Wonder Full, proprietary multimodal models with a one-million-token context and new Active Memory and Active Focus features for long-running agents. Wonder is on Meta's API at $4 per million input tokens; Wonder Full is restricted to Muse Code subscribers.
Meta has released the first models from its Superintelligence Labs, the division led by Alexandr Wang. Muse Wonder and the larger Muse Wonder Full accept text, images, audio and video, output text, and support a context window of about one million tokens. Mark Zuckerberg described them as a much larger and more advanced class of model than the earlier Muse Spark family. Neither ships with downloadable weights, a decisive break from Meta's open Llama era.
Two features are aimed at agents. Active Memory retains relevant personal and professional context across conversations, while Active Focus keeps an agent pursuing its goal through obstacles and failed attempts, revising plans and asking permission before consequential actions.
On the Artificial Analysis Intelligence Index, Wonder Full scores 56, level with Claude Sonnet 5.5 and behind Claude Opus 5.5 on 58, while Wonder scores 52. Meta's own results show Wonder Full ahead on SWE-Marathon, though rivals lead on Terminal-Bench 4.0 and HealthBench Pro. Independent testers, including Abacus.AI's Bindu Reddy, have questioned the frontier framing and promised their own benchmarks.
Access is tiered. Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, with cached input at $0.20. Wonder Full is not on the API at all and is reserved for Muse Code subscribers on $50 and $200 monthly plans. Data handling differs too: API use defaults to no training on submitted content, while Muse Code subscriptions default to a Contributor mode that allows Meta to use submissions for model improvement unless switched off.
Why it matters
Meta's shift from open weights to gated, proprietary frontier models removes the largest corporate sponsor of open AI from the field, leaving Mistral, Chinese labs and smaller players to carry it. For enterprise buyers, the data-handling defaults are a reminder to read the terms of every tier, not just the model card.
Indicops view
Government and regulated-industry clients in India should note the Contributor-mode default carefully; any subscription that can train on submitted content is unsuitable for official data unless explicitly disabled and contractually confirmed. We maintain a checklist of data-use terms across providers and apply it before any new model enters a client environment.
More in Models
Back to the newsroomGoogle launches Gemini 4 Hydrogen, a $1.50-per-million-token model for coding and agents
Google has released Gemini 4 Hydrogen, a faster and cheaper tier of its Gemini 4 family aimed at everyday coding and agentic work. It costs $1.50 per million input tokens and $7.50 per million output tokens across a 1M-token window, and is rolling out through the Gemini API, AI Studio and consumer products.
Alibaba opens a two-week preview of Qwen4-Max with a 1M-token context, but no benchmarks or weights
Alibaba Cloud has opened a time-limited preview of Qwen4-Max, its next flagship model, through the Token Plan only. It offers a 1M-token context and up to 393,216 output tokens, but the documentation lists no pricing, benchmarks or downloadable weights, and the model retires on 25 November.
Anthropic releases Claude Fable 5.2 and Mythos 5.2, 30 per cent faster and 20 per cent cheaper
Anthropic has released Claude Fable 5.2 for general use and Mythos 5.2 for vetted cyber and life-sciences organisations. The company says output is up to 30 per cent faster than the 5.1 generation, at $8 per million input tokens and $40 per million output tokens, with lower rates of safeguard blocks.