Sunday, 28 June 2026

The race between OpenAI and Anthropic just got more interesting, but it's almost beside the point now. Yes, GPT-5.6 is here to challenge Claude Mythos 5 on coding tasks, and yes, these benchmarks matter to some builders. But what's actually reshaping the AI landscape this week isn't the model leaderboard—it's the shift toward agents that can act autonomously in the real world.

Google's move to add Computer Use to Gemini 3.5 Flash is the kind of feature that sounds niche until you think about what it means. An AI that can see your screen and automate your desktop isn't just another capability upgrade. It's a different category of tool. The same goes for what Perplexity is doing with Computer for Counsel, its new legal platform that routes work across twenty-plus models and ties every answer to a source. That's not about model capability anymore—it's about orchestration and workflow. And Microsoft's new Copilot features for Excel, including connectors to third-party data, are clearly chasing the same thread: making AI useful for actual work, not just chat.

But here's where it gets thorny. By the way, Google itself just warned that websites can expose AI agents to hidden traps. If you're building agentic systems that navigate the open web, or if you're considering Coinbase's new AI agent tool that trades crypto on your behalf, you're now operating in territory where the AI isn't just generating text—it's making financial decisions based on what it encounters. That introduces a whole new class of risk. Hidden prompts, adversarial inputs, exploits designed specifically for agents. The security community is going to be playing catch-up here for years.

I find the video generation trend equally significant, if less dramatic. Billions are flowing into AI video startups, and the market is legitimately in its early phase. The industrial capacity here will eventually reshape content creation, advertising, and film production. But unlike agents, which are actively moving money and controlling your systems right now, video generation feels almost quaint by comparison—powerful, yes, but safer because it's generative rather than agentic.

The real question emerging from this week isn't which model wins the next benchmark. It's whether we're moving fast enough on security and oversight to match the pace at which these systems are being deployed into consequential workflows. Microsoft adding oversight features to Copilot is a small acknowledgment of that problem, but I suspect we're underestimating the complexity ahead.