Apple Hands Siri to Google's Models and OpenAI Guts ChatGPT's Interface for Its IPO

01OpenAI's IPO Prep Has a Body Count, and ChatGPT's Interface Is First

OpenAI told the SEC it wants to go public. Inside the company, the consequences are already visible in the product.

The company confirmed it submitted a confidential draft S-1, the registration document a company files before an IPO. OpenAI says it has not set timing for any next step. Confidential submissions carry no public schedule, so the date is not the story. The pressure the filing creates is.

That pressure points at ChatGPT. Ars Technica reported OpenAI is preparing to overhaul the chatbot, recasting it as a route into higher-margin products ahead of a potential IPO. The framing is financial, not technical. A free chatbot that burns compute on every query is a cost center. A funnel that moves users toward paid, higher-margin services is a different line on a prospectus. "Chat is dead," per the reporting, describes a business decision about gross margin, dressed as a redesign.

The logic running through ChatGPT runs through Sam Altman's other ventures too. TechCrunch reported that Tools for Humanity, the identity-verification company behind Altman's eye-scanning project, is struggling to generate revenue and plans to downsize its staff. The report did not specify how many jobs. The timing sits inside the same week as the S-1 disclosure.

Two companies, one orbit, one direction. The profitable parts get pushed forward and the parts that don't generate revenue get cut or reengineered. A public-market filing rewards demonstrated margin and punishes open-ended spending, and that incentive is now reaching products and payrolls across Altman's holdings.

For ChatGPT's users, the redesign decides what stays free and what becomes the paid destination. The free tier was the growth engine that made OpenAI a household name. Treating it as a top-of-funnel cost rather than a product changes who pays and for what. Developers building on the platform inherit whatever pricing structure the margin math requires.

The S-1 itself stays sealed until OpenAI decides to move. What the filing already reveals is operational: a chatbot being rebuilt around margin and a sister company cutting staff in the same stretch of days. The financial restructuring meant to satisfy public-market investors is the news, and it started before any share is sold.

02Apple Built Its Privacy Brand on Doing AI Itself. The New Siri Runs on Google's Models.

For years Apple sold a single idea about artificial intelligence: it would do the hard parts itself, on the device, where no one else could see your data. At WWDC, the company confirmed something it had only been rumored to do. The foundation models behind the next Apple Intelligence were co-developed with Google, using the technology behind the Gemini family.

Apple framed it as a "deep" collaboration that unlocks a "huge upgrade," according to its announcement. The models run both on-device and on Apple's servers through its existing Private Cloud Compute setup, the company says. A new system orchestrator coordinates features across apps and adapts responses to whatever task is open. The pitch is system-wide intelligence with the privacy story intact.

The harder fact sits underneath the architecture diagram. Apple revealed its plans for a smarter, more conversational Siri two years ago, then watched that Siri fail to ship. The version arriving now, which Apple calls "an entirely new version of Siri," exists because the company reached for an outside model after its own efforts stalled. The firm that built its brand on self-reliance is delivering its marquee AI feature on a competitor's foundation.

The tone of the keynote matched the climbdown. TechCrunch reported that this year's demos looked notably grounded, a shift it tied to the $250 million Apple agreed to pay to settle false-advertising claims over earlier AI promises. The settlement followed marketing that described capabilities the product did not deliver. The 2026 keynote, by TechCrunch's account, played less like a launch and more like a spouse reading off a completed chore list.

Apple did not say which devices get the higher-power model with speech generation and improved dictation, leaving the most capable tier undefined. It also did not address the obvious question its own privacy marketing raises: how a model co-built with Google fits a story long told as Apple-alone. The company that spent a decade contrasting itself with data-hungry rivals now ships its central AI feature on their engine.

03A designer who used to throw out every AI mockup now builds prototypes instead of Figma files

The designer had a rule about large language models: every time he reached for one, it let him down. Last year he tried Copilot and Cursor to tweak a game he had built. Neither produced working changes. At an earlier job he used Gemini to outline product briefs and generate wireframes, then discarded all of them.

Each failure shared a trait. He kept aiming the tools at work he was already good at, and they did the job worse than he would have.

That changed after he joined Jane Street last summer. Surrounded by unfamiliar OCaml and Bonsai, he leaned on AI to cover the gaps. The surprise landed elsewhere. It rewrote the part of his job he knew best.

His old design loop ran through spec docs, Figma mockups, written proposals, and implementation reviews with developers. Now he skips most of it. He writes a short description of the problem and his proposed fix, opens his editor, starts a build and a server, and feeds that description to Claude as the prompt.

From there the work becomes iteration on a real prototype, not a picture of one. He gets basic functionality running to prove the idea is possible. He pushes to a development environment and asks users what they think. Then he submits the feature, Jane Street's version of a pull request, behaving exactly the way he wanted. A prototype inside the actual codebase, he writes, has felt better in nearly every way than a mockup.

The pull runs deeper than one designer's habits. A separate GitHub feature request, #65697, asks Anthropic to ship an official Claude Desktop for Linux, and it has collected 522 points on Hacker News while consolidating several earlier filings. The filer notes a concrete blocker: Claude Code plugins are developed and tested against Claude Desktop extensions, which has no Linux build. Plugin work currently forces developers to switch operating systems. Cowork already runs the Claude Code binary inside a Linux VM, so the execution path exists but ships to no published Linux target.