Apple Exploring Ways to Run Much Larger AI Models Directly on iPhones

The most interesting thing about OpenAI's new "super app" isn't the launch itself — it's what the launch tacitly admits. ChatGPT Work, unveiled Thursday, is pitched as an agent that sticks with a project for hours, reaches across your apps and files, and turns a vague goal into finished work. That's a meaningful shift from chat-as-answer-machine to chat-as-colleague, but it's also OpenAI conceding that the chatbot interface alone was never going to be the moat. Anthropic has been quietly winning the enterprise agent conversation for months, and this launch reads as OpenAI's answer to that pressure as much as it does a genuine product leap.

What strikes me is how crowded this space has become in a matter of days. GPT-5.6 arrived promising "more intelligence per token" — a phrase that tells you OpenAI is now competing on efficiency and cost as much as raw capability. Meta Superintelligence Labs pushed out Muse Spark 1.1 the same week. And then there's Lyzr, an enterprise agent-building startup reportedly raising $100 million at a $500 million valuation. None of these are coincidences. Everyone building agent infrastructure right now is racing toward the same conclusion: the value isn't in the model anymore, it's in the model doing something autonomously and reliably enough that a business will trust it with real work.

That's also why Guillermo Rauch's comments about Vercel caught my attention more than the headline launches did. Vercel now sees roughly half of its 6 million daily deployments triggered by coding agents, chewing through more than a trillion tokens. That's not a demo statistic — that's production infrastructure quietly being reshaped by agents that write and ship code without a human in the loop for every step. When the CEO of a major cloud platform says internal automation and coding are the two use cases actually delivering value today, that's a more honest signal than any keynote slide about "ambitious work."

By the way, the Apple story sitting alongside all this agent news is worth pausing on. Reports that Apple has met with PrismML about running much larger models directly on iPhones — no server round-trip — point at a different but related problem: agents are only as useful as they are fast, private, and cheap to run at scale. If OpenAI and Meta are racing to make agents smarter, Apple is racing to make them local. Those are not competing bets so much as complementary ones, and I suspect the winner of this decade's AI race will be whoever solves both simultaneously rather than either one in isolation.

The open question I keep coming back to: when every major lab ships an "agent for your hardest work" within the same week, what actually differentiates them a year from now — the model underneath, or who you already trust with your files?