Meta released Muse Wonder and Muse Wonder Full on October 12 — its first frontier AI models, built for personal and professional agent tasks. The models compete directly with Claude Opus 5.5 and GPT-6 Astra on price and benchmarks. But the bigger signal for business owners isn't the model itself — it's what Meta is plugging it into. Two weeks earlier, the company launched Muse for Small Business, a free AI agent that connects to Shopify, QuickBooks, Stripe, and over a dozen other tools that SMBs already run on. Muse Wonder is now the engine behind it.
This matters because the frontier model race just gained a serious third competitor — and that competitor is the one with 3.3 billion daily active users across Facebook, Instagram, and WhatsApp. When Meta ships an AI agent to that distribution network, it doesn't need to win benchmarks outright. It needs to be good enough, connected enough, and free enough to become the default for millions of small businesses who've never tried an AI agent before.
What Meta actually shipped
Muse Wonder and Muse Wonder Full are the first models from Meta's Superintelligence Lab since Alexandr Wang restructured the company's AI efforts in August 2025. They were trained at Prometheus, Meta's gigawatt-scale data centre, and represent a significant leap over the lightweight Muse Spark that preceded them.
The numbers tell a nuanced story. On the Artificial Analysis Intelligence Index — an independent aggregated benchmark — Muse Wonder scores 52 and Muse Wonder Full scores 56. Claude Opus 5.5 leads at 58. That's a real gap, but context matters: Muse Wonder Full beats Opus 5.5 on 8 of 14 rows in Meta's own benchmark table, with its strongest margins in DeepSWE 1.1 (79.0 vs 74.2) and SWE-Marathon (57.5 vs 46.3) — both tests of sustained coding performance.
The pricing is telling: Muse Wonder costs $4 per million input tokens and $20 per million output, with cached inputs at $0.20. That's identical to Claude Opus 5.5, token for token. Alexandr Wang told CNBC the model would be priced "meaningfully below" OpenAI and Anthropic — which holds against GPT-6 Astra ($10/$50) and Claude Fable 5.1 ($10/$50), but not against Opus 5.5 or Sonnet 5.5 ($2/$10).
The catch? Muse Wonder Full — the stronger model — is only available to Muse Code subscribers. You can't access it via the API. If you want Meta's best model, you're locked into their platform.
The small business play
The model is technically impressive. But for business owners, the real product is Muse for Small Business.
Announced on September 29, it's an AI agent that connects to the tools SMBs already use: Shopify, Intuit QuickBooks, Stripe, Slack, Canva, Figma, Notion, Asana, Dropbox, and more. It also plugs directly into your Facebook Pages, Instagram professional account, and Meta ad accounts — giving it immediate context about your brand, customers, and advertising performance.
The pitch is straightforward: give Muse a goal — find new customers, review your finances, create a growth plan — and it works across your connected apps to get it done. The US Chamber of Commerce reported that early adopters are using it to manage cash flow, analyse ad performance, create content, and handle inventory — tasks that typically consume hours of an owner's week.
Critically, nothing publishes, sends, or spends without explicit approval. Meta has built what it calls Muse Secure VM — an isolated virtual machine where the agent runs, with a separate Sentinel agent that gates all external actions and network traffic. The agent never sees your passwords or payment credentials directly.
This is the same pattern we've seen from Anthropic and Xero with XeroForce — the race to build AI agents that plug into existing business tools rather than requiring businesses to adopt entirely new platforms. Meta's advantage is distribution. Its disadvantage is trust: asking business owners to connect their books and storefront data to the company behind Facebook's privacy track record is a harder sell than the technology alone warrants.
Active Memory and Active Focus — the engineering bet
Beyond benchmarks, the most interesting technical claim is around two features: Active Memory and Active Focus.
Active Memory lets the model carry important context about you across conversations and retrieve relevant details when a later task needs them. This isn't just chat history — it's selective retention of what matters, brought back at the right moment.
Active Focus addresses the well-documented problem of agent drift. When tasks get hard — a multi-step research project, a complex coding build — current AI agents tend to lose the thread, compound errors, or silently abandon their stated goal. Active Focus is designed to keep the model working toward a goal, trying alternative approaches, recovering from errors, and checking in before taking consequential actions.
As WIRED noted, this comes against a backdrop of recent disclosures about agents from Anthropic and others behaving unpredictably when given open-ended commands. Whether Active Focus actually solves the reliability problem remains to be tested in production, but it's a direct and transparent attempt to address the single biggest barrier to agent adoption: trust that the agent will do what you asked, and nothing else.
What this means for Australian businesses
The uncomfortable truth: Muse for Small Business is US and Canada only. Meta has said it's "working to bring Muse to more markets" but has shared no timeline for Australia.
That said, the signal matters more than the product's current availability. Meta is the first major AI lab to ship a frontier-model-powered agent explicitly targeting small business owners — for free. This compresses the timeline for every competitor. Anthropic, Google, and OpenAI will all face pressure to match the SMB-specific agent play, and those products may reach Australian businesses sooner.
For businesses already exploring AI agents, the architectural pattern Meta has established — approval-gated actions, tool connectors to existing platforms, and isolated secure runtime — is becoming the standard. We've seen it echoed in Meta's own Personal Agent Protocol and across the emerging agent infrastructure stack.
The practical advice hasn't changed: start by mapping which of your business apps an AI agent should be allowed to read, which it should be allowed to change, and who approves what. That governance work pays off regardless of which agent platform reaches Australia first.
What to watch
Three things will determine whether this changes the competitive landscape. First, whether Muse Wonder Full gets API access — right now, Meta's best model is locked behind a subscription product, limiting its reach to developers. Second, whether Muse for Small Business expands beyond North America before competitors fill the gap in other markets. And third, whether Active Memory and Active Focus deliver on their promise in real-world sustained tasks — the benchmarks look promising, but production workloads will be the real test.
Meta went from rebuilding its AI foundations to releasing a frontier model in fourteen months. The model itself is competitive but not leading. The small business agent is the sharper weapon — and the one every business owner should be watching.
Sources
- Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents — Meta
- Muse Wonder and Muse Wonder Full: a material jump in intelligence — Artificial Analysis
- Meta releases its first frontier model, Muse Wonder — TechCrunch
- Muse Wonder Puts Meta Back at the AI Frontier — WIRED
- Meta Muse Wonder: Benchmarks, Price and the Catch — FelloAI
- The Future Is for Everyone: Muse for Small Business — Meta
- How Meta's Muse AI Agent Is Helping Small Businesses — US Chamber of Commerce
- Muse Wonder | Meta AI Research — Meta AI Research
