Claude Mythos leaks, AI's 'Second Moment,' and the rise of vertical models

A draft blog post reveals a new Anthropic model tier. Vertical AI products start beating frontier models on their home turf. And political pressure on AI ramps up.

The leak of the week is Mythos, but the bigger trend is the rise of vertical-specific AI that outperforms general models. That changes the strategic playbook for any business with proprietary data.

Anthropic's 'Claude Mythos' leaks

My take: A draft blog post revealed a new Anthropic model tier beyond Opus, being trialed for cybersecurity applications. We don't know the full capabilities yet, but the leak alone tells you Anthropic is pushing well beyond current frontier. If you've planned your AI strategy around current model limitations, those constraints are about to shift again.

AI's 'Second Moment' has arrived

My take: Q1 2026 is being framed as a major inflection point: AI moves from assistants (chatbots) to true agents (autonomous systems with goals). Open Claw was the catalyst. The right way to think about this for your business: stop asking 'how can AI help my team?' and start asking 'what could autonomous agents do for us?' Different question, different answers.

The era of vertical models begins

My take: Intercom's Apex for customer service and Cursor's Composer 2 for coding both now beat frontier models like GPT-5.4 on their specific domains. The takeaway for any vertical SaaS or service business: your proprietary data is more valuable than ever. Fine-tuning a specialized model on your data may now genuinely outperform what the major labs give your competitors.

Zero-employee companies hit $6M ARR

My take: Pulsia, a platform that produces agent-run businesses, reportedly reached $6M annualized with a single founder and no human employees. Whether the model survives is another question, but the proof point matters. The cost floor of operating a business has fundamentally shifted.

Apple's Gemini partnership: bootstrapping on-device AI

My take: Apple can now 'distill' large Gemini models into smaller proprietary models that run on-device. This is the strategy that turns Apple into a credible AI player without having to compete head-on with frontier labs. Worth watching because it positions Apple devices as the default home for personal AI agents.

ARC AGI 3 launched as a new benchmark

My take: Tests an AI's ability to learn and reason by playing simple graphical games with no instructions. Current frontier models score less than 1%. Humans score 100%. Worth remembering when the AGI hype gets ahead of itself: today's models are genuinely brilliant at some things and genuinely terrible at others. Don't extrapolate one to the other.