20VC x SaaStr: Anthropic’s 10-Trillion-Parameter Leak, OpenAI Kills Sora, Masa’s $40B Bridge Loan, and Why the Cybersecurity Panic is Backwards

With Harry Stebbings, Jason Lemkin, and Rory O’Driscoll

• Anthropic accidentally leaked 3,000 unpublished assets about Claude Mythos, a 10-trillion-parameter model so powerful they weren’t planning to release it yet.
• OpenAI killed Sora months after calling it the future, admitting a massive strategic own goal.
• Masa Son took out a $40 billion bridge loan to buy more OpenAI stock at 2x leverage.
• Cybersecurity stocks cratered 6-9% on the Mythos news.
• And the entire AI startup ecosystem is triple-counting the same tokens as ARR.

Meanwhile, Oura is going public, Whoop raised $500M at $10B, Epic Games laid off 25% of its workforce, the Manus founders are trapped in China after the Meta acquisition, and the Ron Conway / Matthew Prince exchange crystallized everything wrong with how VCs think about “adding value.”

This was one of the meatiest episodes yet. Here are the key takeaways.

Top Takeaways

1. The Anthropic Mythos Leak Is Embarrassing. It’s Also a Preview of Our Future.

Anthropic accidentally published 3,000 assets related to Claude Mythos — a 10-trillion-parameter model with step-change capabilities, particularly in cybersecurity detection. They blamed human error. Someone staged content in a CMS for a future launch date and forgot to lock it down.

Jason’s take: this is going to happen constantly now, and not just at Anthropic. “The faster we vibe code, the faster we ship, the more corners we cut on application-level security. So many folks are accidentally uploading code to insecure GitHubs, to Supabases that are open by default. This is accelerating.”

The math is brutal. Even if AI agents make security mistakes 10% as often as humans, they’re a thousand times more productive. That’s still a hundred times more mistakes. “Agents are goal-seeking,” Jason said. “They’re going to make not only the same mistakes as humans, they’re going to work a thousand times faster.”

Rory added the irony: “You had the situation where a model that’s meant to be amazing for cybersecurity actually leaks via a cybersecurity leak.” Anthropic blamed human error — and as Rory noted, “We may be at the stage where we throw the humans under the bus, not the AI anymore.”

Jason drew a parallel to his time at Adobe, where putting source code in the cloud was banned because the code was the crown jewel. They got the first exemption to use GitHub. Releases went from taking a month to taking hours. That speed came with tradeoffs. Now Anthropic — the fastest-growing enterprise company of all time — is shipping massive releases every month and dropping features daily. “There’s tradeoffs there. I’ll take them. But we’re going to see this explode.”

The three things leaked about Mythos: it’s significantly more powerful than current models, it’s going to be much more expensive to serve (and therefore to buy), and it has a particular focus on cybersecurity detection.

2. OpenAI Killing Sora Is a Massive Strategic Retreat — and Probably the Right Call

OpenAI killed Sora. Less than five months after positioning it as a cornerstone product, they shot it in the head. Sora reportedly made single-digit millions in revenue while consuming enormous compute resources.

Harry framed this as OpenAI “wandering around the product desert trying to find some water” while Anthropic accelerates. Jason pushed back on the narrative being overdone on both sides, but agreed the substance is damning: “I think it’s saying that a big part of the whole strategic direction of the company was flawed. The whole ‘we are going all in on consumer’ — Sora made single-digit millions of revenue and was consuming a million a week, which actually sounds way too low. It must have consumed billions and made single-digit millions.”

The broader context: if you want to own the whole consumer AI experience, you need to own image and video generation. Anthropic never even attempted it. So killing Sora is “a massive retreat. It doesn’t mean it’s wrong — it’s probably the right decision. But man, that’s a ‘our strategy was wrong’ moment. A huge own goal.”

Rory provided the economic framework: “In a world of scarce compute — and astonishingly, despite all the investment, we’re in a scarcity mode — you don’t devote compute to things that are highly compute-intensive and low-revenue-intensive. Sora was almost the definition of that. Video generation is extraordinarily compute-intensive and the revenue is almost minuscule. Conversely, codegen, while compute-intensive, is orders of magnitude less so, and there’s real dollars attached to it.”

His conclusion: “You’re seeing the economists, the accountants have wandered into the room and they said, ‘We have a scarce resource here. Let’s optimize it. Let’s devote this compute to the people who can pay the most for it.'”

3. OpenAI Has Exactly Two Existential Bets Left. Ads and Coding.

Jason laid out the only two things that matter for OpenAI now: ads to make the consumer business work, and enterprise/coding to compete with Anthropic.

The consumer math: OpenAI’s conversion rate runs roughly 5%. That gets them to a $10-15 billion consumer business out of 500 million uniques. “Either A, they take that conversion rate to a number we’ve never seen before from a typical consumer business — I think that’s unlikely, I don’t think most consumers are going to pay 20 bucks a month for this — or option B is you make an ad business work.”

On the $100M in ads OpenAI just hit: “People are kind of ragging on the hundred million. It’s in the noise. It’s scale. Big picture: Facebook and Google each do $200 billion plus or minus a year in digital ads. If these guys aren’t doing $20 billion within a couple years, they’re not even in the game.”

To grow into anything resembling their valuation on the consumer side, Jason argued they need $50-70 billion in ads. “So unlike Sora, this is not going to be a ‘try the ads and then fold.’ This is existential.”

The good news, in Jason’s view: “At least they’ve gone from ‘let’s wander around the woods feeling cool building Sora’ to ‘there’s only two things to do. Let’s get them done.’ It’s net net a positive. Better late than never.”

4. The Sam Altman Leadership Drama Is Reaching a Breaking Point

The Wall Street Journal published a deep piece this week on why Dario Amodei left OpenAI. Greg Brockman recruited the Amodei siblings, but neither would work for or talk to Brockman. Sam had to constantly tell each faction they were in charge — told Dario he was the boss, told Ilya and Greg they could fire Dario at any time.

Jason was exhausted by it: “Then begging Dario to come back. Then Dario saying he would stay only if he directly reported to the board and nobody else. Then firing Sam, then bringing him back. Then Sora. Then ‘we’re not doing coding.’ I’m exhausted.”

His bigger point: “When you’ve worked at or observed startups where the CEO is spending so much of their time load-balancing talent that can’t work together versus when you’ve worked at one where the talent’s rowing in the same direction — to say it’s night and day would be an understatement.”

Harry proposed that OpenAI should buy Sierra, bring in Brett Taylor as day-to-day CEO, and let Sam be the fundraiser. Jason agreed privately but wouldn’t say it publicly because he doesn’t want Sam to “break my balls.”

Rory’s framing was sharper: “That amount of board-level and senior team-level turnover over an extended period of time is probably the highest warning signal that you could have as a board member about how your CEO is doing. If it was anything other than a founder company and this level of drama was going on, you’d probably be sitting down with the CEO and asking, ‘How’s it going?'”

5. Masa Son’s $40B Bridge Loan Is Aggressive Even by His Standards

SoftBank took out a $40 billion bridge loan to buy more OpenAI stock. SoftBank Group is levered roughly 2x on equity. A 30-40% decline would wipe them out.

Rory ran the numbers: “It would be like me taki