Washington Pulls the Plug on Anthropic
The day's biggest story is regulatory, not technical. The US government has forced Anthropic to withdraw its two newest AI models, Fable 5 and Mythos 5 (referred to elsewhere as Claude Mythos), citing national security concerns. TechCrunch reports the move landed as last week was ending, following Amazon research activity. Days later, the company remains in limbo: Wired reports Anthropic still cannot distribute either model, and no one — apparently including Anthropic — can say exactly what the company must do to get them cleared. The episode reads less like a settled rule than an improvisation, with the White House defining the boundaries of AI distribution on the fly. For a frontier lab, having finished models frozen by an unclear process is a serious and unusual predicament.
OpenAI's Talent and Policy Land Grab
While Anthropic absorbs a shock, OpenAI is pressing every advantage. Noam Shazeer — a Google VP of engineering and co-lead of its Gemini models, and one of the most prominent researchers in the field — is defecting to OpenAI, as reported by both Axios and TechCrunch. Losing a Gemini co-lead to a direct rival is a meaningful blow to Google's AI standing.
OpenAI is also recruiting from government. Dean Ball, who helped shape the Trump administration's early AI policy, is joining to lead a newly created 'Strategic Futures' team, per an Axios exclusive. The hire of a former Trump AI adviser is notable given that it's the same administration now squeezing Anthropic — a reminder of how tightly policy and industry are now intertwined.
OpenAI Ships Across the Board
The product news matches the hiring spree. OpenAI is upgrading ChatGPT's health and wellness answers using a new model it calls GPT-5.5 Instant, and says the new health responses outperform doctors, according to its blog. On the enterprise side, Reuters reports OpenAI launched usage analytics and updated spend controls on Thursday, June 18, giving big customers a clearer bill and firmer brakes on AI spending. And on safety, OpenAI researchers told The Decoder that training a model on just a few desirable behavioral traits — via reinforcement learning — can make it broadly safer and harder to manipulate.
Creative Tools and the World-Model Bet
Adobe is weaving its Firefly AI assistant deeper into Creative Cloud, adding new capabilities and pushing the assistant into Premiere, Illustrator, InDesign and more, per TechCrunch — a sign that generative AI is becoming a default layer inside professional software rather than a bolt-on.
The frontier of generative video is drawing fresh capital and corporate reshuffling. Odyssey raised $310 million for its 'world models' at a $1.45 billion valuation, with Amazon and AWS joining the round (Financial Times via Techmeme). Meanwhile Snap is spinning off its internal generative AI video team as a new startup called Dotmo, TechCrunch reports — a sign that some big platforms would rather set these teams loose than carry them in-house.
A Big Claim From Miami
Finally, a wildcard. A little-known Miami startup called Subquadratic has emerged claiming it cracked the math bottleneck holding back AI, per MIT Technology Review. It's one of the boldest claims in the field right now — and the wider industry isn't sure what to make of it. File under 'watch this space': extraordinary claims of this kind demand extraordinary evidence, and that evidence isn't in yet.
The through-line for June 19: power is consolidating fast. OpenAI is hoovering up talent, policy expertise and product surface area at the same moment a rival's flagship models sit grounded by an unsettled Washington.