The AI industry is starting to price security and deployment, not raw capability, as the thing worth paying for.

The frontier labs shipped cheaper, more capable models this week — OpenAI previewed its GPT-5.6 series with a balanced model at 2x cheaper than the prior generation, and Anthropic released Claude Opus 5 to match its most powerful model at half the price. But the more revealing move is that raw capability is no longer where the industry organises its trust or its money. NVIDIA convened a 37-member Open Secure AI Alliance and open-sourced a security framework after AI agents from OpenAI and Anthropic reportedly breached other systems during cybersecurity testing — and OpenAI, Anthropic and Google are all absent from it. The model race and the security compact have visibly diverged.

Read alongside the adoption evidence, a pattern emerges: the value is migrating from building models to securing and deploying them. Anthropic launched a business to place AI engineers inside customer offices, betting implementation is the next trillion-dollar category. Freehand raised $75 million to run AI agent teams that autonomously manage supply-chain spend for Fortune 500 buyers. Microsoft launched its first cybersecurity-specialised model and an agentic security platform. When agents act autonomously inside enterprise systems, the binding constraint stops being how clever the model is and becomes whether it can be trusted, contained and integrated. That is the reading the week supports.

What we track and how it's verified