The ClickedOn AI Wrap: Week of 13 Apr 2026

On Wednesday, Meta Superintelligence Labs launched Muse Spark, its first proprietary model. Built over nine months by a team led by former Scale AI CEO Alexandr Wang (poached in a $14 billion deal), the model represents a ground-up overhaul of Meta's AI strategy.

The numbers are solid. Muse Spark scores 52 on the Artificial Analysis Intelligence Index, putting it fourth globally behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. On HealthBench Hard, it topped every rival with a 42.8% score. It accepts text, image, and voice inputs, and introduces "Contemplating mode," which orchestrates multiple sub-agents reasoning in parallel on complex tasks.

What makes this notable for marketers: Muse Spark is rolling out across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta AI glasses in coming weeks. That's 3+ billion users getting access to a frontier-class model embedded directly in the apps they already use daily. If your customers interact with Meta properties, they'll soon have a capable AI assistant sitting inside every conversation.

The strategic shift matters too. Meta moved from open-source Llama models to a closed, proprietary system. That signals Meta sees AI not as a public good to distribute, but as a competitive moat to protect. For businesses that built workflows on Llama's open weights, this could mean rethinking dependencies.

Anthropic made a rare move this week: announcing a model it explicitly will not make publicly available. Claude Mythos, the company's most powerful system to date, has already uncovered thousands of zero-day vulnerabilities in every major operating system and web browser. Rather than release it, Anthropic launched Project Glasswing, a cybersecurity coalition backed by $100 million in usage credits and $4 million in direct donations to open-source security organisations.

The partner list reads like a who's who of tech: AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, Microsoft, and Nvidia all have access to Claude Mythos Preview for defensive security work. More than 40 additional organisations that build or maintain critical software are also involved.

For business leaders, this is a significant signal. Anthropic is effectively telling the market: "Our model could break your software, so we're helping you patch it first." If your organisation runs any major platform, the vulnerabilities being found right now may directly affect your infrastructure. Keep an eye on the security patches rolling out over the coming months, many will trace back to Glasswing findings.

Three of AI's fiercest competitors started sharing intelligence this week through the Frontier Model Forum. The trigger: Anthropic's February disclosure that 16 million Claude exchanges across roughly 24,000 fraudulent accounts were attributed to Chinese labs MiniMax, Moonshot/Kimi, and DeepSeek conducting unauthorised distillation.

The economics explain why this matters. Training a frontier model costs roughly $1 billion. A successful distillation run costs as little as $100k to $200k. That 5,000:1 cost ratio made individual defences ineffective, forcing cooperation.

The labs are now exchanging four specific data types: fraudulent account fingerprints, proxy infrastructure data, hardened signup flows, and chain-of-thought elicitation classifiers. It's the first time the Forum has functioned as an active threat-intelligence operation against a specific external adversary.

For Australian businesses using AI APIs, this may mean tighter account verification and rate limiting in coming months. If you're building products on these platforms, factor in potential disruptions to API access patterns as the labs roll out new anti-distillation measures.