AI ACTUALLY
Issue No. 26 — Sunday, July 12, 2026
This week, in order: a hardware company sued its own AI hardware partner for allegedly stealing its trade secrets, two other AI companies got dramatically cheaper on the same day, and Anthropic spent its energy building you a dashboard that asks if you’ve been overdoing it. Read that last one twice. Coffee first. Then this.
Apple just accused OpenAI of running a corporate espionage ring
Apple sued OpenAI on Friday in federal court, and the filing does not read like a polite disagreement. Apple alleges that OpenAI’s hardware chief, a former Apple VP who now runs hardware for OpenAI — directed Apple employees interviewing at OpenAI to bring “actual parts” to their interviews. Batteries. Logic boards. The stuff you’re supposed to leave in the building.
Apple also names a former Apple engineer, who allegedly kept his work laptop after leaving for OpenAI, found a bug that let him back into Apple’s internal cloud storage, and downloaded a stack of confidential files anyway. Separately, Apple claims OpenAI misled one of Apple’s own suppliers into demonstrating a proprietary metal-finishing technique, on the understanding that Apple had signed off. It hadn’t.
The backdrop makes this stranger: Apple and OpenAI are current partners — ChatGPT is built into the iPhone’s operating system. OpenAI also bought Jony Ive’s hardware startup, io Products, for $6.5 billion last year, and is reportedly racing to ship its first consumer device sometime this year. Apple’s lawsuit argues that device is now built on stolen foundations. OpenAI’s response, in full: it has “no interest in other companies’ trade secrets.”
Why it matters: Companies accuse each other of stealing ideas constantly; they rarely accuse each other of coaching employees on how to sneak parts out the door. Apple is functionally saying OpenAI’s hardware ambitions were built by mining Apple’s own staff for spare parts and manufacturing know-how. Whatever OpenAI ships this year, it now ships with a lawsuit attached to its origin story.
Meanwhile, OpenAI had a very big product week anyway
Lawsuits aside, OpenAI actually shipped something Thursday: GPT-5.6, its new model family (Sol, Terra, and Luna), is now fully public — no more waitlist, no more “20 pre-approved partners.” That’s worth a beat of its own, since three weeks ago this exact model was rationed to a short list of trusted organizations at the government’s request. It’s now just... available.
Alongside it, OpenAI folded its Codex coding tool into a rebuilt ChatGPT desktop app under the name ChatGPT Work — a workspace that can browse, edit files, run for hours on multi-step projects, and generally do the thing Anthropic’s Claude Cowork already does. Sol lands just under Claude Fable 5 on general intelligence, but beats it on agentic coding, runs faster, and costs roughly a third less — Sol’s top pricing is $5/$30 per million tokens, versus Fable’s steeper rate.
Then Meta piled on the same week: Muse Spark 1.1 went live on a public API at $1.25/$4.25 per million tokens — about a quarter of what frontier labs charge. Mark Zuckerberg framed it as a direct shot at what he called the “very extreme” margins at rival labs.
Why it matters: Two different companies just made the same bet in the same week — that “nearly as good, much cheaper, no usage anxiety” beats “best, but rationed and pricier.” Anthropic’s models are still the benchmark everyone measures against. They’re no longer the benchmark for what things cost.
While two rivals built bigger empires, Anthropic built you a mirror
No new model. No price cut. No superapp. What Anthropic shipped this week was Claude Reflect — a private dashboard that looks back at your own Claude usage over the past 1, 3, 6, or 12 months and shows you your own patterns: what you actually use it for, when, and how often. It adds quiet-hours settings, break reminders, and self-reflection prompts, mapped to something Anthropic calls its “4D AI Fluency Framework.”
Nobody asked for this the way people ask for a cheaper model. That’s rather the point.
Why it matters: Every other AI company this month has been trying to become a bigger part of your day — more tabs, more tokens, more tasks handed off. Anthropic’s move was to build a tool whose entire function is asking whether that’s actually going well for you. It’s a strange thing to ship in a week when your competitors are racing to eat more of your attention, and it’s exactly the kind of thing that only makes sense if you’re not trying to win the same race.
Safe to ignore this week
TeraWulf’s $19B Anthropic lease, again — we covered this Tuesday. It hasn’t grown a new number since.
Illinois’s AI safety law — same law, same story, one newsletter cycle later.
Elon Musk says he won’t “cut off” Anthropic — a promise with a shelf life we’re not tracking.
Anthropic adds Ben Bernanke to its oversight trust — a governance move, not a product one.
China weighing limits on who can use its own AI models — real, but a one-source rumor that needs another week to firm up.
Grok 4.5 launches — benchmarks look strong. So did the last four launches this month.
A Google chatbot flaw let one rogue agent read other users’ conversations — patched in June, disclosed this week. Worth knowing your vendor patched it; not worth a section.