On October 12, 2026, Meta released Muse Wonder and Muse Wonder Full, the first frontier-class models from Meta Superintelligence Labs (MSL). The company calls them the most capable models it has built, designed for coding, long-running autonomous agents, and complex personal and professional tasks. The launch marks Meta's serious entry into the frontier model race, backed by gigawatt-scale training infrastructure the company calls Prometheus.
What the two models are
The release has two members:
- Muse Wonder: the standard variant, available now to developers through the Meta Model API (as
muse-wonder-1.0) and inside Muse Code, Meta's developer-facing coding agent. API pricing is $4 per million input tokens and $20 per million output tokens, with cached inputs reported at $0.20. - Muse Wonder Full: the larger variant, running at maximum reasoning for the hardest, longest-running jobs. It is not served through the API at launch — it is gated to Muse Code subscribers on the $50/month Power Usage and $200/month Maximum Usage tiers.
Both models carry a 1 million token context window, putting Meta in the same bracket as Google's Gemini 4 Argon. That context length matters specifically for agentic work: tasks that span dozens of tool calls, long codebases, or extended multi-step plans degrade quickly once a model runs out of room to hold prior steps in memory.
Active Memory and Active Focus: the differentiators Meta is pitching
Meta built the launch story around two features. Active Memory is designed to carry relevant information across conversations — remembering what matters to a user over days, weeks, and months without requiring the same context to be resupplied. Active Focus is Meta's name for persistence: trying other approaches when a task stalls, recovering from errors, asking for clarification, and checking in before anything consequential.
Meta says the models reach expert level in law, finance, and health alongside frontier coding performance, and claims Muse Wonder outperformed rival systems on MemoryBench, a test of continual learning and memory construction. A note of caution: Meta published no benchmark figures at launch, and a fuller safety evaluation report was promised rather than released — so the performance claims can't yet be checked independently. Because agents read messages, calendars, and documents, Meta says the models were trained not to share private information without permission and to refuse help with harmful acts.
Meta AI chief Alexandr Wang called them "true frontier models" aimed at "personal and professional agent tasks." Stock market reaction was positive, with Meta shares gaining roughly 4% on the day, per Reuters.
Background: Meta's 14-month AI overhaul
The Wonder release is the product of an AI overhaul Meta began roughly fourteen months ago, building on Muse Spark, the rebuilt model series Meta introduced in April 2026 (the same "Muse" family name this assistant runs on). Meta is pitching Wonder squarely at agent work — navigating an insurance claim, parsing a lease, sorting out someone's finances, or building a working app — the same territory OpenAI and Anthropic have staked out in 2026.
Availability is being staged: Muse Wonder will soon power experiences in Muse, Meta's personal AI agent, rolling out to subscribers first. The Muse app itself remains limited to North America for now, with no announced date for Europe — a reminder that regulation still gates where agents can be sold.
Why it matters for founders and investors
Two things stand out. First, memory is now a first-class product feature at the model layer, which changes what agent startups should build: products that manage persistent user state are no longer defensible as thin wrappers — but products that own the workflow and data around that memory still are. Second, Meta is now competing directly with developer-tool startups via Muse Code; any founder building AI coding products just gained a well-funded, distribution-rich competitor with 1M-token context and $50/month pricing. For investors, the question to ask every agent startup this quarter: what do you do that the model itself doesn't, and won't, by next quarter?
Angeled.com connects founders with investors.
Sources: Meta newsroom/press release (Oct 12, 2026); Reuters reporting on the launch; PYMNTS; Notebookcheck; kie.ai and MIT Sloan India summaries of model specifications and availability.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



