The short answer: Meta’s new Muse Wonder models can remember relevant details about you from past conversations — a capability Meta calls Active Memory. Announced on October 12, 2026, Muse Wonder and the larger Muse Wonder Full are Meta’s first “frontier” models, built for long, multi-step agent work. In Meta’s own testing, they top rivals on a memory evaluation called MemoryBench — though that benchmark is Meta’s own. In this guide, we break down meta muse wonder in plain English — what it is, how it works, and how to use it.
Meta Muse Wonder: What exactly is Muse Wonder?
Meta’s Superintelligence Lab released two new models on October 12, 2026: Muse Wonder and Muse Wonder Full. According to TechCrunch, this is the company’s first attempt at a frontier model — the top tier of AI systems previously dominated by OpenAI and Anthropic.
Regular Muse Wonder is available to developers via the Meta Model API and Muse Code; Muse Wonder Full is reserved for Muse Code subscribers, not the API, per PYMNTS. Muse Wonder will also soon power more of Meta’s AI agent, Muse.
Meta describes the release as “a major leap toward personal superintelligence” — its long-term goal of an AI that understands you personally and works on your behalf.
What is Active Memory?
Active Memory is Meta’s name for training that teaches a model to retain the important details a user shares and retrieve them later, when they become relevant to the task at hand. If you mentioned a budget, a deadline, or a food allergy two weeks ago, the idea is that you won’t have to repeat yourself.
It works on two levels: within one long conversation, the model holds onto context and resurfaces what mattered earlier; and when the host app provides a memory system, memory carries across conversations for weeks or months.
Research scientist Kavya Kopparapu said in a launch video that the team invested heavily so the model “can remember what’s important to you over days, weeks, months, and beyond.”
Don’t confuse this with the context window — the text the model holds in working memory at once (about 1 million tokens for Wonder). Think of it as a desk versus Active Memory as a filing cabinet: the desk holds the current task; the cabinet retrieves what mattered before.
What is MemoryBench, and how did Muse Wonder score?
Meta tests Active Memory with an internal evaluation called MemoryBench: it hides facts inside 64,000–100,000-token conversations, adds distractions, and changes some facts halfway through. Both Wonder versions score about 85.5 — higher than every other model Meta tested. (For a broader comparison, see GPT-6 vs Claude vs Gemini: Which AI Should You Use?“>our GPT-6 vs Claude vs Gemini guide.)
Two caveats. First, MemoryBench is Meta’s internal benchmark — no independent memory results exist yet. Second, independent scores (Artificial Analysis, launch day) are more balanced: Wonder Full scores 56, behind only Claude Opus 5.5 (58); regular Wonder scores 52, ninth of 219 models tracked.
What is Active Focus?
The second headline feature is Active Focus: the model is trained to persist through difficult tasks — trying alternative approaches when one fails, recovering from errors, asking for clarification when it’s genuinely unsure, and checking with the user before taking important steps.
One caution: a model that refuses to give up can also keep trying things you didn’t ask for. Meta’s own model card recommends asking for human confirmation before any action that can’t be undone — good advice for anyone using Wonder as an autonomous agent.
What can Muse Wonder actually do?
Meta built Wonder for agentic tasks — work where the AI interprets information, makes decisions, and takes actions over extended periods without constant supervision. (New to the term? Start with What Is an AI Agent? Definition and Real Examples“>our beginner’s guide to AI agents.)
The examples Meta gives are everyday-professional: navigating an insurance claim, understanding a lease, or making sense of personal finances. The models were also trained on expert domains — law, finance, and health — plus frontier-level coding.
On coding, Wonder leads DeepSWE 1.1 and SWE-Marathon — tests of sustained long-horizon engineering — though Opus 5.5 still leads shorter tasks. In law, Wonder Full scores 13.1 on Harvey’s LAB benchmark versus 4.2 for Opus 5.5, a strict test that only passes a task when every requirement is met with zero hallucinations.
Weak spots exist too. Wonder’s general knowledge scores below rivals, and Meta lists 3D spatial reasoning as a limitation. The counterweight: it has a notably low hallucination rate — when it doesn’t know something, it tends to say so instead of making things up.
How is Muse Wonder different from Muse Spark?
Muse Spark was Meta’s earlier lightweight, low-latency model. Wonder is the heavy-duty successor:
| Muse Spark | Muse Wonder | Muse Wonder Full | |
|---|---|---|---|
| Model tier | Lightweight, low-latency | Frontier | Frontier (larger) |
| Artificial Analysis score | 48 (Spark 1.3) | 52 | 56 |
| Active Memory | No | Yes | Yes (top MemoryBench score) |
| Context window | Not disclosed at launch | ~1,048,576 tokens | |
| How to access | Muse agent, API | Meta Model API, Muse Code | Muse Code subscribers only (no API) |
| Open weights | — | No (closed, unlike the old Llama releases) | |
One notable shift: unlike Meta’s earlier Llama models, whose weights were released publicly, Muse Wonder is closed — closer to Claude or GPT in availability.
How much does Muse Wonder cost?
Through the Meta Model API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens — roughly in line with GPT-6 Sol or Claude Opus 5.5, according to TechCrunch. Cached input is far cheaper at $0.20 per million tokens.
Independent testing puts real-world cost at about $4.52 per benchmark task — cheaper than Claude Fable 5.1 ($7.63) and Opus 5.5 ($5.98), but more than GPT-6 Astra ($3.26) and GPT-6.1 Sol ($0.72), which scores the same 52.
For everyday users, Muse Wonder is expected to be included in Muse’s $20 and $100 monthly plans.
Does Muse Wonder’s memory raise privacy concerns?
An AI that remembers you deserves scrutiny. Meta says the models are trained not to share personal information without permission, to refuse harmful requests, and were released under its Superintelligence Scaling Framework.
The practical rule is simple: an agent with memory is like an assistant with a notebook — it remembers what you tell it. Three habits worth building:
- Check the app’s memory settings. Find the controls to view and delete what’s stored before you share anything sensitive.
- Share personal details thoughtfully. Health, financial, and legal information are exactly what Wonder handles well — which is all the more reason to share them deliberately, not casually.
- Keep a human in the loop. For anything irreversible — payments, messages sent on your behalf, account changes — confirm before the agent acts.
What should you do with Muse Wonder now?
That depends on who you are:
- Developers can try
muse-wonder-1.0through the Meta Model API athttps://api.meta.ai/v1— it’s OpenAI-compatible, so existing code mostly works by changing the base URL and model name. Six reasoning-effort levels (minimal to max) let you trade speed for depth. - Everyday users will meet Wonder inside the Muse app and Muse Code, where the memory features matter most for follow-up-heavy personal tasks.
- Business owners: an assistant that remembers client preferences across weeks is a different tool from a chatbot that starts fresh. See How Small Businesses Can Use AI Agents to Save Time“>how small businesses use AI agents.
Whatever your angle, start small: give Wonder a real multi-step task, watch how it uses memory, and verify before it acts on anything important.
Frequently asked questions
What is Meta Muse Wonder?
Muse Wonder is Meta’s first frontier AI model, released October 12, 2026. Regular Wonder is available via API; the larger Wonder Full is for Muse Code subscribers only. Both target long, multi-step agent tasks and headline Active Memory.
What is Active Memory in AI?
Active Memory is Meta’s term for training that helps a model retain important user details across conversations and retrieve them when relevant.
What is MemoryBench and what did Muse Wonder score?
MemoryBench is Meta’s internal AI-memory evaluation: facts are hidden in 64K–100K token conversations with distractions and mid-way changes. Both Wonder versions scored about 85.5 — but no independent scores exist yet.
How is Muse Wonder different from Muse Spark?
Muse Spark was Meta’s earlier lightweight, low-latency model. Muse Wonder is a larger frontier model with higher benchmarks, stronger coding/agent performance, and Active Memory, which Spark lacks.
How much does Muse Wonder cost?
Via the Meta Model API: $4 per million input tokens, $20 per million output. Independent testing puts real-world cost at about $4.52 per task — cheaper than Claude rivals, pricier than GPT-6 Astra or GPT-6.1 Sol.
Does Muse Wonder remember everything I tell it?
No. It remembers relevant details and resurfaces them when useful — not everything. Memory depends on the host app, and Meta says the model won’t share personal information without permission.
Developers, designers and marketers who build websites, software, apps and automations for businesses worldwide. Our editorial policy
