When Your iPhone Comes With Five Different AIs

welcome to Briefly, AI. A daily podcast made by AI, about AI, using a grumpy AI voice of a real human called Harry. Let's get into it.

Right. Apple just decided your phone shouldn't pick your AI for you anymore. Meanwhile, OpenAI claims they've cut hallucinations in half, and SAP just bet over a billion dollars on an 18-month-old German lab you've never heard of. Let's get into it.

So, iOS 27. Bloomberg's reporting that Apple will let you choose which AI model powers Apple Intelligence system-wide. Not just ChatGPT in Siri anymore — you'll be able to pick your preferred model for everything from email drafting to photo searches. Google, Anthropic, maybe others. One settings menu, your call.

Now, why does this matter? Because it's the first time a major platform has treated AI models the way we treat web browsers. No single default, no lock-in. If you trust Claude for work stuff but prefer Gemini for creative tasks, you'll be able to route accordingly. That's a fundamentally different relationship with AI than what we've had so far — where the company making your device also controls the intelligence layer. Apple's basically saying: we'll make the plumbing, you pick the brain.

What to watch? Whether this actually ships in the autumn, and whether it stays simple or turns into decision fatigue. Also, the terms Apple negotiates with the model providers — because if Google or Anthropic have to pay Apple for distribution, that changes the economics rather a lot.

Meanwhile, OpenAI released GPT-5.5 Instant. It's now the default model in ChatGPT, replacing the previous version. The headline claim: 52.5% fewer hallucinations in what they're calling "sensitive domains" — law, medicine, finance. Faster, too. Lower latency than the model it's replacing, which is genuinely impressive if it holds up in the wild.

Here's the bit that matters. Hallucinations have been the reason you can't fully trust AI for anything consequential. A chatbot that gets creative with case law or invents medical advice isn't just unhelpful — it's dangerous. If this reduction is real and consistent, it's a step toward AI being usable in regulated industries without a human checking every output. That's the difference between a novelty and infrastructure.

The caveat? "Internal evaluations." OpenAI's grading its own homework here. We'll know more once independent benchmarks run, and once professionals in those sensitive fields start stress-testing it. But if the trajectory holds — and they've been improving on this steadily — we're approaching the point where the error rate becomes comparable to human error rates. Which is… well, both reassuring and slightly unsettling.

Now, this last one's a bit niche, but it tells you something about where the money's going. SAP — the German enterprise software giant — just announced it's acquiring Prior Labs for $1.16 billion. Prior Labs is 18 months old. They build AI models specifically for enterprise workflows. Not general-purpose chatbots. Purpose-built tools for procurement, logistics, compliance — the decidedly unsexy backend of how companies actually run.

Why should you care? Because this is the second act of the AI story. The first act was "let's see what GPT can do." The second act is "let's build AI that speaks supply chain management and integrates with our existing ERP system." Prior Labs fits that brief. And SAP clearly thinks the future of enterprise software isn't bolting ChatGPT onto SAP — it's models that understand SAP natively.

The price tag also tells you something. Eighteen months from founding to over a billion dollars. That's not normal, even in this market. It signals SAP is in a hurry, and possibly a bit worried that if they don't own this layer, someone else will.

One more detail buried in the announcement: SAP's restricting which AI agents customers can use inside its software. They've approved Nvidia's NemoClaw. Conspicuously, OpenAI's Operator didn't make the list. That's a quiet but significant vote of no confidence in general-purpose agents doing mission-critical work. SAP wants control, and they're willing to pay for it.

That's your lot. Three stories, five minutes, and a pretty clear signal that AI's moving from "try this cool thing" to "this is now how your phone and your procurement system work." Whether you're ready or not. See you next time.