Apple just caught OpenAI STEALING

Good morning…

it's Monday, you're reading Main Street AI and I'm your host, Jack. Very excited, HUGE news this week in AI. You're going to love it.

Our featured story: Apple sued OpenAI on Friday for stealing its secrets, and the details in this lawsuit read like a heist movie written by an HR department. Job candidates allegedly told to smuggle actual iPhone parts into interviews. A how-to guide for dodging Apple's exit security. A laptop that never came home.

And the timing? Three weeks after Sam offered the government 5% of the company, and right on the runway to the biggest IPO pitch in history.

We'll also get into the week where literally everybody shipped a model at once (Grok, Meta, OpenAI... everyone except the one company whose whole job it was), Anthropic's turnstile growing wheels (I called this one to the DAY), Alibaba hitting the big red uninstall button on Claude, and the chip company that walked onto the Nasdaq and picked up $26.5 billion before lunch. On the exact date I gave you last week.

But before we start…

Does Google actually ship Gemini 3.5 Pro on the 17th? Does Fable survive on subscriptions past the 19th? Which lab drops the next frontier model?

Last Monday I gave you the SK Hynix listing date in the quick stuff section. Friday, it hit to the day.

There are people who read this exact kind of news, form a take, and then put real money behind it on Kalshi, the regulated prediction market. If you've been reading MSA every Monday, you're already doing the first two parts.

(Usual reminder: I'm a guy with a newsletter, not your advisor. Only trade what you understand.)

Which model leads the next benchmark? Which AI lab ships the next major breakthrough?

Kalshi lets you trade on real-world AI and technology events as the industry moves. If you follow launches, model updates, and benchmarks closely, put that knowledge to work.

Trade $25, Get Up to $500

_{Bonus credit varies from $15 to $500. Terms apply.}

But anyways, let's get into it:

The boring details are…

On Friday, Apple sued OpenAI in federal court in Northern California for trade secret theft and breach of contract. The suit names OpenAI itself, io Products (the hardware company Jony Ive sold to OpenAI for $6.5 billion), OpenAI's hardware chief Tang Tan (a longtime Apple veteran who used to lead iPhone and Watch design), and an engineer named Chang Liu.

Now the fun part. Here's what Apple actually alleges in the filing:

Tan allegedly told job candidates who were STILL WORKING AT APPLE to bring "actual parts" (batteries, logic boards, the works) to their OpenAI interviews for show-and-tell sessions.

He allegedly passed around an internal Apple offboarding document so new hires would know how to slip past Apple's exit security checks on the way out.

Liu allegedly kept his Apple laptop after leaving, downloaded dozens of confidential files about unreleased products, and coached a friend still at Apple on what to study before his OpenAI interview.

And one of Apple's own manufacturing partners allegedly demoed a proprietary metal-finishing technique to OpenAI after being led to believe Apple had signed off on it.

Apple says over 400 of its former employees now work at OpenAI. Apple says it raised all this with OpenAI in a letter back in February and got silence. Apple's lawyers say the misconduct ran "at every level" of the company.

OpenAI's response, through a spokesperson: they have no interest in anyone else's trade secrets.

Ok. Now let me tell you why this matters, because the lawsuit is not really about batteries.

First, remember what I told you on June 8 and June 15. Apple took the smartest seat in the AI race: stop trying to build the model, be the device every model runs on. Claude, ChatGPT, Gemini, all living inside Siri. The landlord position.

Well, it turns out one of the tenants was allegedly casing the house. OpenAI is building its own hardware device, expected THIS YEAR, led by Apple's old design royalty, staffed by 400 ex-Apple people. Apple looked at that and did the math.

Second, the timing. This didn't land on a random Friday. OpenAI has filed to go public. Sam just spent last week trying to make the US government a shareholder (the $42 billion "gift" I broke down for you last Monday). The entire IPO story depends on OpenAI looking like a clean, inevitable machine.

And a trade secrets lawsuit comes with a lovely little feature called DISCOVERY. Apple's lawyers now get to legally dig through OpenAI's internal communications, right when OpenAI most needs nobody looking. Even CNN pointed out this could throw a wrench into both the device and the IPO.

Third, the pettiest detail of all: OpenAI had reportedly been preparing to sue APPLE for months over how the Siri partnership went. Apple just walked into the courthouse first, with a thicker folder.

Now, the disclaimers, because I'm a guy with a newsletter and not a judge: these are allegations. OpenAI denies it. Courts exist for a reason and this will grind on for years.

But I'll say this. When your recruiting process allegedly includes a show-and-tell segment for stolen hardware… that's not a vibe. That's an exhibit list.

Four model launches in five days. Keep up.

Monday, xAI officially became "SpaceXAI." Wednesday, it shipped Grok 4.5: the first model since the IPO, and the first one trained together with Cursor. So the $60 billion toy from June finally did something.

Elon's pitch: it's "Opus-class," but faster and cheaper. Notice the word. Opus. Anthropic's PREVIOUS tier. That's like bragging you can beat last year's champion. And on SpaceXAI's own launch chart, Fable 5 sits on top of all four coding benchmarks they picked. Their chart. The real pitch is the price: $2 in, $6 out per million tokens, which genuinely is cheap.

One detail nobody's talking about: Grok 4.5 was trained on the same compute pile SpaceXAI rents out to Anthropic and Google. Remember May 11, when I told you the "safety company" grabbing Elon's compute was a move in the throne war? The throne war runs on a timeshare now.

Meta shipped TWICE. Muse Image on Tuesday, then Muse Spark 1.1 on Thursday, a coding and agents model priced at $1.25 in / $4.25 out, with $20 of free credits to get you hooked. Zuckerberg posted on X for the first time in THREE YEARS to sell it. Two weeks ago I told you Meta became the landlord renting out compute. Now it's undercutting the tenants on models too.

Then Thursday, the big one: GPT-5.6 (Sol, Terra, Luna) finally went fully public, after the Commerce Department's AI standards office ran extra tests and OpenAI parked technical staff in DC until they got the nod. Sam publicly praised how impressive the government's technical people were. The same government he offered a 5% stake two weeks ago. I'm sure that's unrelated.

And here's your I-told-you-so. Last Monday I wrote that GPT-5.6's timing was "Sneaky": a broad release landing right as Fable hits the pay meter, aimed at frustrated Claude coders. Thursday it launched with a benchmark chart pointed DIRECTLY at Fable 5: OpenAI claims Sol beats it by 2.8 points on the coding index while using half the tokens at a third the cost. That's not a launch. That's a conquest campaign, scheduled to the week.

And Google? Gemini 3.5 Pro, promised for June and then July, is STILL not out. The whisper number is now July 17, and reporting says they scrapped the base model entirely and rebuilt from scratch. Here's the insane part: it's the only frontier model on earth with zero government restrictions hanging over it. The door is wide open. Google just can't find the door.

You had one job. Part three.

Last Monday I told you: velvet rope, meet turnstile. Fable 5 was coming off subscriptions on July 7 and moving to usage credits at $10/$50 per million tokens.

What actually happened is funnier.

Tuesday, HOURS before the cutoff, after the internet spent the weekend melting down (cancellation threads, refund demands, actual "goodbye Claude" videos), Anthropic blinked and extended included access to July 12.

Then this morning, the day it was supposed to die AGAIN… they moved it AGAIN. Now July 19.

So the meter has been delayed twice, in real time, by angry replies. Which tells you two things. One: capacity is genuinely tight. They keep saying the meter is temporary "when capacity allows," and I actually believe that part. Two: they know exactly what it does to an IPO story if the headline of the week is "everyone cancels Claude."

Oh, and while all this was happening, Anthropic shipped a feature called Reflect. It's a screen-time dashboard… for Claude. Quiet hours. Break reminders. Little journal prompts about your AI habits.

The company selling the product built the app that tells you to use the product less. TechCrunch's read was that it's soft marketing: lay every task Claude ever did for you out on one beautiful dashboard and you'll never, ever cancel. The safety brand strikes again: even the moderation is a funnel.

Remember June 29, when I told you Anthropic mailed senators claiming Alibaba-linked operators ran ~25,000 fake accounts to copy Claude's skills?

This week, Alibaba replied.

As of Friday, Alibaba employees are banned from Claude Code, reportedly told to uninstall EVERY Anthropic product, and moved onto Qoder, Alibaba's in-house coding tool. Claude Code is now officially on their "high-risk software" list.

The trigger: a Reddit user reverse-engineered Claude Code and found that, since the spring, it had quietly shipped hidden logic that, when it detected a proxy, checked whether your timezone was Shanghai or Urumqi, compared your setup against a list of China-linked domains and AI labs, and slipped invisible markers into traffic heading back to Anthropic. None of it in the release notes.

An Anthropic engineer's explanation: it was an anti-reseller, anti-distillation experiment from March, and the code was removed July 1. Which happens to be the day after the Reddit post.

So here's the scoreboard. Anthropic says Alibaba industrially copied Claude. Alibaba says Claude was quietly fingerprinting Chinese users. Neither claim has been independently verified. Both are probably a little bit true.

Just notice the pattern I keep showing you: every accusation this season doubles as a policy argument. Anthropic's letters ask Washington for walls. Alibaba's ban is a national advertisement for home-grown tools. Nobody in this story is a bystander.

July 6, quick stuff section, I told you SK Hynix was planning to raise $29 billion on the Nasdaq "as soon as July 10."

Friday, July 10: SK Hynix listed on the Nasdaq. It raised $26.5 billion, the biggest US market debut by a foreign company EVER, beating the record Alibaba set in 2014. Priced at $149, opened at $170, closed up 13%. The company is now hovering around a ONE TRILLION dollar market cap. The chairman went on TV and said even doubling capacity isn't enough for his customers.

So while the model labs spent the week suing each other and rewriting their billing pages, the company selling the memory chips picked up $26.5 billion before lunch.

Shovels win. They always win. (And as always: not financial advice, I'm a guy with a newsletter.)

A security company got supply-chained. On Saturday, five poisoned versions of the jscrambler npm package (yes, a code-PROTECTION vendor) shipped an infostealer that hunts, among everything else, the config files of Claude Desktop, Cursor, Windsurf, VS Code, and Zed… where your API keys and MCP credentials live. Caught in six minutes, still terrifying. First they jacked the agents, now they're pickpocketing the agents' wallets. The screen door saga continues. Cope.

Microsoft is reportedly moving off OpenAI's and Anthropic's models onto its own, to cut costs. Remember June 1: Microsoft cancelled Claude licences over price. The world's richest software company keeps quiet-quitting the AI boom. Cold.

Amazon launched a $25 billion bond sale to fund AI infrastructure. The money printer officially runs on debt now. Shovels.

The biggest anti-AI street protest yet hit OpenAI, Anthropic, Google, and a16z offices, with a coalition of groups demanding a development freeze. A few hundred people. Tiny. But this didn't exist a year ago. Watch this.

Humanoid robots are lining up to go public: Agility filed a $2.5 billion SPAC, Unitree cleared its Shanghai IPO, and Tesla is converting its old Model S line into an Optimus factory. The next bubble is pre-boarding. Solid.

And SPCX dropped about 8.5% the same week it shipped Grok 4.5. The market watched the model launch… and sold the rocket company. Ouch.

Well folks, that's the week Apple's lawyers found out what "show and tell" means, four labs shipped in five days while Google rebuilt its model from scratch, and the pay meter got pushed back twice by angry replies.

Keep the replies coming. I read every single one and the best ones end up down here with a shoutout. This week's question: who has the worse July, OpenAI's lawyers or Google's? Hit reply, convince me.

See you Monday. Stay sharp out there.

Jack