Welcome back to the Technology newsletter 🤖
OpenAI quietly swapped out the brain inside ChatGPT for a new default model that hallucinates a lot less and remembers more about you.
Meanwhile Apple is reportedly about to break its ChatGPT exclusive and let iPhone users pick which AI runs Siri, and Anthropic put its CEO on stage next to Jamie Dimon to pitch banks on a new top-end Claude.
Let's get into it 👇
🍏 According to a Bloomberg report Tuesday, Apple is planning a feature in iOS 27 (due this fall) that will let iPhone users choose which AI service powers Apple Intelligence (including Siri, Writing Tools, and Image Playground). They expect users could choose between Google's Gemini, Anthropic's Claude, and the existing ChatGPT integration. Internally, Apple calls the feature "Extensions," and it'll live in the Settings app. iOS 27 gets its first big reveal at WWDC on June 8. Apple’s renewed focus on hardware looks a lot like a quiet admission that it can't win the AI race alone. The Settings app is where pride goes out with a whimper.
💰 The AI boom is starting to look a lot less like a Nvidia monopoly. AMD reported earnings Tuesday afternoon and clobbered every estimate: $10.25 billion in revenue (up 36%), data-center sales up 57% to $5.8 billion, and forward guidance well ahead of expectations. The stock ripped 16% higher on Wednesday. AMD also said its first full rack-scale AI system, "Helios," ships later this year, and that both OpenAI and Meta have already signed up for it. Each one of those rack systems sells for upward of $3 million. For years, Nvidia was the only game in town. AMD just kicked the door open.
🤖 Anthropic just announced two major updates at once: a new top-end model called Claude Opus 4.7, and a set of ten "finance agents" designed to do the slog work that bank analysts currently grind on, like building pitchbooks, running KYC checks, and closing the books at month-end. Anthropic says Claude is already in production at JPMorgan, Goldman Sachs, Citi, AIG, and Visa. The analysts running these models may soon find the model running without them.
🇳🇴 Norway is set to formally sign onto Pax Silica, a U.S.-led coalition of countries trying to build a non-China-dependent supply chain for chips, AI, and critical minerals. That makes Norway the 15th member, joining the U.K., Japan, South Korea, Australia, India, Greece, and others. The State Department is announcing the deal this week. But the secret sauce for this that Norway happens to run the largest sovereign-wealth fund on Earth: about $1.7 trillion. And it sits on real reserves of the rare-earth minerals modern chips need.
If you've used ChatGPT at any point in the last 48 hours, you've been talking to a different AI than you were a week ago.
On Tuesday, OpenAI quietly replaced the default model that runs ChatGPT for everyone (free, Plus, and Pro users alike) with a new one called GPT-5.5 Instant.
There was no big keynote, no Sam Altman livestream, no "this changes everything" tweet, just a small note in the help docs and a fresh model picker.
But the change is pretty significant, and it's worth understanding why.
The headline number from OpenAI's own internal testing: GPT-5.5 Instant produces 52.5% fewer hallucinated claims than the model it replaced when asked about high-stakes topics like medicine, law, and finance.
"Hallucinating" is the polite industry term for "confidently making things up,” the thing that's caused ChatGPT to invent fake court cases, recommend nonexistent drugs, and tell people their flight is on time when it isn't.
A 52% drop is the kind of number that, if it holds up in the wild, makes the difference between a chatbot you fact-check and a chatbot you can start to trust a bit more on a first pass.
The new model also scored an 81 on the AIME math test (a closely watched benchmark), up from 65 on the old default (a bigger leap than usually happens in a single point release).
Plus and Pro subscribers get a second feature too, and this one may make bigger waves in your day-to-day: ChatGPT can now search your past conversations, your uploaded files, and your Gmail to give you more personalized answers.
Ask it "what was that recipe you helped me with last month" or "draft a follow-up email to the contractor I was talking to in March" and it can actually go look.
To address the obvious privacy concern, OpenAI also added a "memory sources" panel that shows you exactly which past chats, files, or emails fed into a given response, with the ability to delete or correct anything you don't want it remembering. So you can still delete that email you sent to your employer with your W9 attached.
Critics will reasonably point out that handing ChatGPT your Gmail is still a bigger trust ask than most people realize.
The bigger picture is that the AI race in 2026 is increasingly being fought not at the absolute frontier (the giant flagship model launches that get keynotes) but in the boring middle, where billions of people interact with the default.
OpenAI competes with Google's Gemini, Anthropic's Claude, and (depending on the week) Elon Musk's xAI by making the version of ChatGPT that the average free user gets handed a little bit better, a little bit less wrong, and a little bit more personal.
GPT-5.5 Instant is exactly that kind of upgrade, and it shipped with a whisper. The AI race isn't being won at keynotes anymore. It's being won at the default, quietly, on a Tuesday while you weren't looking.
What did you think of today's newsletter?