The Barriers are Down: Inside the Race for the Public Markets and Frontier AI

Yesterday in AI | Wednesday, June 10, 2026

The Barriers are Down: Inside the Race for the Public Markets and Frontier AI

The barriers are officially down for the most capable AI model in history, but that's just the tip of the iceberg in a massive week for the AI industry. Today's episode breaks down Anthropic's commercial launch of Claude Fable 5, the security safeguards protecting its zero-day vulnerability capabilities, and the financial strategy behind the timing.
 
Plus, we unpack the unprecedented race to the public markets as both OpenAI and Anthropic gear up for concurrent IPO filings. We look at the multi-billion-dollar cash burn projections that public investors are being asked to fund, and how a surprise emergency injunction from European regulators just shifted the entire playing field for AI platform distribution. Finally, we explore Taiwan’s historic legal move to criminalize AI chip smuggling and a massive new $25 million open-source genomic partnership designed to solve the biological data bottleneck.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Wednesday, June 10th, and we finally have official confirmation that two of the most powerful AI companies on the planet are locked in a race toward public offerings, while the most capable AI model in history just removed the barriers and let everyone into the club. Let's get into it. Let's start with the story I've been watching for months. Back in April, Anthropic ran something called Project Glasswing, a restricted program that quietly gave access to a model called Mythos to about 40 organizations, not regular users, not even regular enterprise customers. 40. Vetted partners doing defensive cybersecurity work, government agencies, national labs, financial regulators. The reason for the secrecy? Mythos found thousands of security vulnerabilities, including zero days using novel approaches. Zero days are security flaws that even the software's own developers don't know about yet. No patch, no warning, just an open door for attackers. The NSA was using it. The Fed was using it. It could compress weeks of hacker reconnaissance into hours. Tuesday, that chapter ended. Anthropic launched Claude Fable 5, a mythos class model, to enterprise customers and paid subscribers. Anyone with a credit card and a business use case can now access capabilities that two months ago required a background check. The price tag is $50 per million output tokens. Tokens are roughly the chunks of text the AI generates, so think of it as paying per word at scale. For context, that's two times what Opus 4.8, Anthropic's previous flagship model, runs. Through June 22nd, it's included at no extra cost in Pro, Max, and Team Plans. After that, it moves to usage credits. Anthropic's pitch? A smarter model delivers higher ROI, so paying double actually saves you money. Maybe. The math depends entirely on what you're using it for. What's notable about this launch is what Anthropic built around it. New safeguards block certain responses in cybersecurity and biology, the two areas where Mythos was doing the most sensitive work under Glasswing. So the model that found zero days in Firefox is now commercially available, but with guardrails specifically designed to prevent the worst-case uses. Whether those guardrails hold up is a question the security research community will be answering for months. The broader signal here: Anthropic is monetizing its most capable model right as it's heading toward a public offering. The timing is deliberate. Anthropic isn't the only AI giant dusting off its books for Wall Street. Its fiercest competitor actually beat them to the SEC's doorstep. Monday, OpenAI publicly confirmed it had confidentially filed its own S1 with the SEC, and S1 is the formal registration document a company submits when it's planning to go public. Same move, same playbook. There's a twist though. OpenAI actually submitted that filing on May 22nd, about 10 days before Anthropic's June 1st filing, so Anthropic followed OpenAI to the SEC, not the other way around. Both companies kept it quiet until last week. OpenAI has 900 million plus weekly active users. Goldman Sachs and Morgan Stanley are running the deal. They're targeting a Q4 2026 listing. They're also preparing a tender offer, essentially a deal that lets employees cash out their shares before the company officially hits the stock exchange. At the current $852 billion valuation. Given how long some of these people have been waiting, that matters. Here's the part worth paying attention to. That's not a typo. And they've said publicly that the company won't outspend its own revenue for at least four more years. So what they're asking public market investors to buy is a company that's spending at an extraordinary rate, generating massive revenue but not yet profitable, on the belief that whoever owns the most capable AI infrastructure wins everything. Sam Altman said this week that OpenAI went public with the filing news because they expected a leak, which is probably true, but it also lands in the same week as Anthropic's filing and a few months after SpaceX's IPO, three of the most consequential companies in tech hitting public markets simultaneously. Whoever goes first captures more of the scarce capital chasing this sector. What the IPOs actually change. Once these companies are public, they have to disclose a lot more. Revenue trajectory, cost structure, safety incident reporting, the financials behind AI labs have been opaque since day one. That's about to change. While American AI giants prepare to capture public market capital, European regulators are moving aggressively to ensure those same companies can't block smaller competitors from capturing users. Here's the background. Meta has a product called WhatsApp Business API. It lets developers and businesses build AI assistants that live inside WhatsApp. Last October, Meta changed the access terms. Third-party AI chatbots suddenly had to pay a fee. Meta's own AI assistant? Still free. A handful of smaller AI companies, including the maker of Poke.com, a French startup called Agentik, and a Spanish competitor, filed complaints with EU antitrust regulators. Tuesday, those regulators moved. They issued an emergency interim order. Meta has five business days to restore equal access to the WhatsApp business API under the same terms that applied before the October change, and the order can remain in force all the way through June 2029 while the investigation continues. If Meta is ultimately found to have breached EU antitrust rules, the fine is up to 10% of global annual revenue. Why this matters beyond just Meta. The EU just demonstrated that it's willing to use emergency injunctions in AI markets, not wait years for final decisions. These markets move fast. Regulators historically have not. What happened Tuesday suggests that's changing. There's also a structural point here. WhatsApp has 3 billion users. If you're building an AI assistant for consumers or businesses, distribution through WhatsApp is massive. Whoever controls API access controls a major on-ramp to those users. The EU is essentially saying you can't use platform ownership to tilt that playing field toward your own AI. Controlling the software platform is one way to choke out the competition, but the ultimate choke point remains the hardware itself. Taiwan is drafting legislation that would make unauthorized AI chip exports to China a criminal offense. For the first time. Currently, if a company illegally routes NVIDIA chips to a Chinese buyer through Taiwan, the charge might be something like document falsification. Under the new framework, the chip export itself would be criminal, and the scope is broad, not just blacklisted entities, but any Chinese customer for chips above a certain compute threshold and for AI servers containing NVIDIA processors. Taiwan's government is coordinating this with the US. That detail matters because the U.S. closed the Singapore export loophole back in May. That was the pathway where Chinese-owned companies could buy chips by routing purchases through Singapore registered entities. Taiwan closing their own gap means the two most important nodes in the global chip supply chain are now moving in lockstep. The downstream effect, China's ability to stockpile the specialized chips needed to train frontier AI models, the most capable AI systems that exist today, depends increasingly on gray market channels. Criminal liability changes the risk math for everyone in that supply chain. Brokers, logistics companies, local distributors. It's the difference between a regulatory fine and someone going to prison. Finally, while geopolitics threatens to stall frontier AI development by cutting off hardware, a massive new alliance is quietly unlocking the data needed to fuel the next frontier of biological science. Google DeepMind and Google.org announced a five-year partnership with the Welcome Sanger Institute, one of the world's leading genomics research centers, to build what they're calling AI-ready genomic data sets. Five million dollars per year for five years. The datasets will be shared broadly. Here's why this matters, and it requires a tiny bit of context. The bottleneck in AI for biology is often not the model, it's the training data. Alpha Fold, DeepMinds AI that cracked the decades-old problem of predicting how proteins fold into 3D shapes, changed what was possible in biology. But genomics, the study of how fold genomes work, how genes vary, and how they drive disease, has a different problem. The data is fragmented, inconsistently structured, and wasn't built with AI training in mind. You can have the best AI design in the world and still train it badly if the underlying data is a mess. What this consortium is trying to do is generate data sets specifically designed to train AI, building new data from scratch with AI training as the explicit goal, rather than repurposing existing research data that wasn't built for this. The Welcome Sanger Institute sequences more human genomes than almost anyone. If they start generating training-specific genomics data at scale and share it openly, that could meaningfully accelerate what's possible in personalized medicine, drug discovery, and understanding genetic disease. We covered the Chan-Zuckerberg Biohub Protein Structure Database back in May. That was about the structure of individual proteins. This is a layer above about how genomes work, vary, and interact. Different scale, different application. And the open sharing commitment makes this one worth watching. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterday andai.news, or you can find me on LinkedIn, X or BlueSky. And if you like this podcast and want to see it continue, please take a minute to rate and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.