New · October 12, 2026
Muse Wonder
Meta just shipped its first frontier model. Wonder and its bigger sibling Wonder Full replace the lightweight Spark with models built for agentic work — and a memory that spans months, not minutes.
What changed
🧠 Active Memory
Remembers what matters across days, weeks, months — user preferences, ongoing projects, decisions. Meta says it tops rival models on MemoryBench, though raw numbers aren't independently verified yet.
🎯 Active Focus
Keeps multi-step agents from drifting: rethinks strategy when stuck, corrects its own errors, asks follow-up questions instead of guessing. Built for the Muse agent, not just chat.
💰 The price of frontier
$4/M input, $20/M output — identical to Opus 5.5. Meta's open-weights era is over: Wonder Full is locked to Muse Code subscribers ($50–$200/mo), no API, no open weights.
How it scores
Artificial Analysis Intelligence Index (independent): Wonder Full 56, Wonder 52 — vs Opus 5.5 at 58, GPT-6 Astra at 53. Meta's own card shows Wonder Full ahead on some agent/coding tests (57.5% SWE-Marathon vs 46.3% Opus 5.5). One honest note: Meta's safety report hasn't been published yet, so safety claims can't be checked.
5 prompts that use the new brain
Written for Wonder's strengths: long memory, multi-step work, course correction.
Build a living project memory
I'm starting a kitchen renovation. Remember: budget $25k, I hate open shelving, my contractor is Luis (555-0142). Every time I mention the renovation, use these facts without me repeating them — and tell me when a decision conflicts with what you remember.
The anti-drift research task
Research the 5 best CRM tools for a 10-person sales team. If you get stuck on pricing pages or conflicting reviews, change your approach and tell me what you tried. Check in with me before your final ranking.
Month-long habit coach
Help me train for a 10K run over 8 weeks. Remember my progress each week, adjust the plan based on how my runs felt, and call me out if I skip two sessions in a row. Start by asking about my current fitness level.
Contract archaeology
I uploaded three versions of a vendor contract from the last year. Track what changed between versions, flag any term that got worse for me, and remember these terms for future contract questions.
Self-correcting trip plan
Plan a 5-day Tokyo trip under $2,500. If a flight or hotel you found becomes unavailable, find the next best option on your own and tell me what changed. Ask before spending anything.
Good to know
- 1M-token context window, shared between input and output — very long inputs leave less room for answers.
- Both models will power the Muse agent going forward, including on iPad.
- Privacy note: unverified online claims say Muse reads private messages; Meta denies it. Treat as unconfirmed until the safety report lands.
- New here? Set up Muse in 15 minutes → · Get 1B free tokens →
← Muse for iPad · Where Muse is blocked · Muse Spark Open · Muse for Windows → · All news
Not affiliated with Meta. Figures: Artificial Analysis, TechCrunch, Meta model card, Oct 2026. Pricing via Meta Model API docs.