Back In the Game

$10 billion — the walk-away fee SpaceX agreed to pay Cursor this week if Elon doesn't exercise a $60 billion call option to acquire the company by next April. That walk-away is larger than Cursor's entire valuation a year ago, and sits 15% above the $52B round Cursor was already raising from a16z and Nvidia. Elon didn't invest. He bought twelve months of guaranteed exclusivity against OpenAI — who tried to buy Cursor last year and got rejected. Welcome to the week the AI race stopped being a coronation. Everybody's back in the game.

Switch image tooling to GPT Image 2 by Friday. If you've been running Google's Nano Banana, Midjourney, or the older OpenAI models for in-brand imagery, test Image 2 this afternoon. It swept all seven Image Arena categories with a 1,512 score and a 242-point lead over the next model — the first clean sweep in leaderboard history. The text rendering and designed-layout control are the specific gap Nano Banana was supposed to close; it just got wider. OpenArt integrated Image 2 on launch day. Spend an hour, then decide.

Audit your MCP-based agent stack this week. Bloomberg reported Tuesday that unauthorized users bypassed Anthropic's access controls on Mythos — the model Anthropic itself called too dangerous to ship publicly — and that a private Discord has had access since launch. Separately, MCP has a disclosed design flaw exposing millions of AI apps to remote code execution. You don't need to know what's been done with either. You have to assume something has. Pull the list of what's connected, what it can reach, and what happens if the protocol underneath it leaks.

Write down your vertical's moat today. One page. What about your business can the frontier labs service but not eat? Regulated workflows. Physical-world handoffs. Customer relationships measured in decades. Proprietary data no lab can license into a training set. If your answer is "we're the best at prompt engineering in our space," your moat is gone by September. And if the honest answer is "we're too small to be on anyone's roadmap," write that down too. Being beneath the labs' minimum economic deal size is the new moat.

Today's actions touched on vertical defensibility in the AI era and agent security posture — two of the questions we spend much of our time thinking about. If you're staring at these actions three and not sure where to start, that's the conversation we're built for.

Outsider Labs Advisory →

Wednesday, April 22

1. The walk-away fee is bigger than the company was. SpaceX bought the right to acquire Cursor for $60 billion — or pay $10 billion if Elon walks. Cursor was worth $9.9 billion twelve months ago and was raising this week at $52 billion from a16z and Nvidia. The $10B isn't an investment. It's a one-year block fee against every lab trying to eat Cursor, starting with the OpenAI offer Cursor rejected last year. Compute-for-equity with a call option attached is a financial primitive the industry didn't have a name for yesterday. It's about to be the template for the next fifty deals. (Full analysis below.)

2. Anthropic's 40-organization whitelist never held. Bloomberg confirmed Tuesday that unauthorized users bypassed access controls on Mythos — the model Anthropic said it couldn't release publicly because the capability was too dangerous. A private Discord has had access since launch. CISA, the U.S. government's own cybersecurity agency, still doesn't. On April 17 we wrote that Anthropic was "managing a weapon." They weren't. The voluntary-lab-controls framing just failed on the most dangerous product the industry has shipped, and it failed quietly for weeks. (Full analysis below.)

3. GPT Image 2 didn't win the leaderboard. It lapped the field. 1,512 score. 242-point lead. All seven Image Arena categories swept — unprecedented in the history of the benchmark. OpenArt integrated it on launch day. The readable text and designed-layout capabilities are the ones Adobe, Canva, and Google's Nano Banana were hoping to keep. Gone. If Spud — GPT 5.5, per Aligned News — drops Thursday as rumored, OpenAI will ship two frontier products in one week after eight weeks of commentary writing its obituary. (Full analysis below.)

4. Tim Cook's $4 trillion farewell. Cook steps down September 1 after fifteen years and a 10x market-cap run from $350 billion to just over $4 trillion. John Ternus, head of hardware engineering, takes over. Johny Srouji expands his silicon authority. Apple's AI chief left five days ago. Siri is being rebuilt on Google's Gemini. There is no Chief AI Officer. Apple's bet: precision engineering outpaces model-frontier velocity. It's a plausible thesis. It's also the first time since 2011 that Apple's AI strategy doesn't have Cook's name on it.

5. A Chinese humanoid ran a half-marathon faster than any human in history. Lightning — five-foot-five, bright red, 5K at a time — crossed the finish line in under 51 minutes. Jacob Kiplimo's human world record is 57:20. China shipped over 1,000 humanoid robots in 2025. No American company shipped more than 500. UBTech plans 10,000 units this year. The Pentagon just asked Congress for $75 billion for drones — more than most nations' entire military budgets. The automation conversation stopped being about software this week.

Back In the Game

SpaceX this week agreed to pay Cursor $10 billion if Elon doesn't buy the company — a walk-away number larger than Cursor's entire valuation twelve months ago. The structure: a call option to acquire Cursor for $60 billion by next April, or the break fee if SpaceX passes. Cursor was already raising this week at $52 billion from a16z and Nvidia. Elon's number sits fifteen percent above a round already on the table. That's not an investment. That's a block.

Why the structure exists is the worst-kept secret in AI. Cursor's Composer 2 beat Opus 4.6 on Terminal-Bench at one-tenth the price in March. But every frontier coding lab — OpenAI's Codex, Anthropic's Claude Code, Google's Gemini CLI — owns the compute Cursor needs to train the next version. Cursor was renting GPUs from every company trying to kill it. Colossus, at 230,000 GPUs in Memphis today and a million by year-end, was the only neutral-ish exit. SpaceX also ships Grok Code, so neutral is doing heavy lifting — but xAI losing the coding race to Cursor beats Cursor losing to OpenAI, who tried to buy the company last year and got rejected.

The last six weeks belonged to Anthropic. Claude Design shipped and Figma dropped 7% on announcement day. Opus 4.7 ranked #1 on the APEX-Agents investment banking leaderboard. The consensus said Anthropic was running away with the category. This weekend three players came back hard. SpaceX wrote the $10B block. OpenAI swept all seven Image Arena categories with GPT Image 2 — 242 points clear, the first clean sweep in leaderboard history. GPT 5.5, codenamed Spud, is rumored to drop Thursday. Sergey Brin un-retired from Google specifically to "out-code Anthropic."

And while the rest of the industry was climbing back in, the lab that lectures Washington on AI safety lost control of its most dangerous model. Bloomberg reported Tuesday that unauthorized users bypassed Anthropic's access controls on Mythos. A private Discord has had access since launch. On April 17 we wrote that Anthropic was "managing a weapon" through the 40-organization Glasswing whitelist. We were wrong. The whitelist was theater. Sam Altman called the whole Mythos rollout "fear-based marketing." He wasn't entirely wrong.

Every week the industry map gets redrawn and every week it's wrong. What's actually happening is a quiet bifurcation of strategy. xAI is consolidating compute and selling exclusivity to products that need it — Cursor this week, something else next. Anthropic is going vertical inside the enterprise and racing toward an $800 billion IPO. OpenAI is running three plays at once: sweep the image leaderboard, ship the next frontier model three days later, turn on CPC ads inside ChatGPT, land Codex at four million weekly users. Google, after watching all three eat its lunch, pulled Brin out of retirement.

The irony underneath is the compute-competitor paradox. Figma paid Anthropic for compute while Anthropic shipped the product that took 7% off Figma's market cap the day it launched. Cursor paid every lab trying to eat it. The only known exit, as Cursor just demonstrated, is to sell a twelve-month call option on yourself to the one man still outside the game.

Which raises the real question. What are these companies trying to build? Coders will pay — and pay highly — for every marginal point of capability. Consumers will not, which is why OpenAI is turning on ads and Meta is hooking keystroke-trackers to employee laptops to scrape AI training data off its own staff. Enterprise pays slowly and is now openly wary of vendors that turn into competitors. Each lab is picking a different customer to marry. The founder reading this needs to figure out which lab is about to marry their business — and decide what to do about it before the vows.

The frontier went horizontal this weekend. Your defensibility goes vertical. The labs can service your business, but they can only eat it if what you do is something a model replicates at scale. Regulated workflows. Customer relationships measured in decades. Proprietary data no lab can license. Physical-world handoffs. Those are the verticals worth defending. And if you're small enough, you're invisible — Figma got eaten because Figma was big. Being small used to be a disadvantage. This week it became a moat.

At COAI today: Full coverage of the SpaceX-Cursor call option, the Mythos breach, OpenAI's image sweep, and the week Anthropic's safety framing cracked — at getcoai.com.

— Harry and Anthony

Sources: