Trump keeps kneecapping the U.S.’s most promising AI models 

Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week via email here. Bullies vs. Brains: Anthropic’s new scrape with the Trump administration Anthropic is in another fight with the Trump administration, this time over its new Mythos-class models, released last week. On Friday, administration officials panicked after reports that Amazon researchers had tricked Claude Fable 5 into providing cybersecurity information Anthropic had tried to block. Officials gave the company 90 minutes to voluntarily take Claude Fable 5 and Claude Mythos 5 offline. Anthropic, waiting for evidence that the models had actually been compromised, did not immediately comply. The administration then declared the models a cybersecurity risk and barred foreign nationals from using them. Because Anthropic had no practical way to limit access to U.S. citizens only, it shut the models down for everyone. As a result, nobody can use Claude Fable 5 or Claude Mythos 5, likely the most powerful publicly available AI models in the world. That may keep bad actors from exploiting them, but it also prevents cyberdefense researchers and software companies from using them to stop cyberattacks. Even foreign nationals who work at Anthropic are now barred from using the models. The evidence for the panic appears thin. The cybersecurity expert Katie Moussouris reviewed the Amazon researchers’ report and told The Atlantic that Fable had refused a direct request to review insecure code for security flaws, but complied when asked to “fix this code,” followed by additional manual steps. The White House had also reportedly heard that a China-linked group had gained access to Mythos, but the government presented no evidence, and Anthropic disputed the claim. This marks the second major dustup between the Trump administration and Anthropic, and the first was just as silly. The Department of Defense banned the use of Anthropic models by defense contractors big and small after Anthropic stuck firm to its rules barring use of the technology for targeting autonomous weapons and surveilling U.S. citizens. Some feared that after being branded a “supply chain risk,” Anthropic would see its business nosedive. Instead, the good vibes around Anthropic increased and its success with enterprises accelerated. Now an IPO is on the horizon. That the administration would single out an AI company it branded “woke” for punishment is hardly surprising. But the Mythos ban also exposes a deeper problem: The administration is applying a special, ad hoc standard to Anthropic’s models that no other U.S. model faces. That is partly because, under “AI czar” David Sacks, the administration has never established a clear risk framework for evaluating frontier models like Mythos. The result is perverse: Some of America’s most powerful AI models have been taken out of action, while competing Chinese models remain unencumbered by U.S. regulators and continue to improve. To be sure, Anthropic deserves serious scrutiny. But the White House fight over Fable and Mythos has blurred more important questions about how the company actually governs its models. In the Fable 5 release notes, Anthropic said it would begin retaining user prompts for 30 days to help detect misuse, and the data would not be used for training. But the policy still raises obvious privacy concerns. The company also said it would silently degrade Fable 5’s output if it detected that the model was being used to train competitors. That lack of transparency was troubling, and Anthropic has since halted the practice. To skeptics, this was all proof that Anthropic is not in fact so different from its rival, OpenAI. Ben Thompson, the tech analyst who runs the Stratechery publication and podcast network, argued that Anthropic’s willingness to degrade the output of its models to thwart rival model makers points to something darker: that Anthropic “does not think that anyone else other than them should even be making frontier LLMs.” But isn’t it possible that Anthropic really was just trying to prevent other model makers from ripping off its own model’s intelligence, even if it chose a sneaky and ill-advised way of doing it? It’s quite a leap to gather from that that Anthropic believes it’s the only company capable of building frontier large language models. Thompson suggests that silently dumbing down model output is akin to training its models not to assist users, even defense users, in targeting autonomous weapons or surveilling U.S. citizens. “What this degradation represented was both the capability and willingness of Anthropic to silently alter its models to achieve its policy preferences,” he writes. “In other words, Anthropic willfully validated some of its critics’ worst fears in terms of being a supply chain risk.” But doesn’t Anthropic have a legitimate interest in setting limits on how its own models are used, even in defense? For example, the company explained that it prohibited the use of its models for domestic mass surveillance in part because current U.S. law “does not yet account for the breadth and speed of mass surveillance that AI could enable.” Why SpaceX is buying Cursor just days after its IPO Like Mark Zuckerberg, Elon Musk badly wants to have a horse in the race for the biggest and best general-purpose AI model. Right now Meta and xAI (now part of SpaceX) aren’t competitive with Anthropic, OpenAI, and Google Deepmind, but they want to build or buy the pieces they need to compete. An AI-assisted coding tool was a big missing piece for xAI, hence SpaceX’s acquisition of Cursor and its people. And computer code is a key ingredient for training AI models, so Cursor’s coding data could end up increasing the quality of xAI’s Grok models and chatbot. xAI has been codeveloping models with Cursor since April, a partnership that gave SpaceX the option to buy the startup for 60 billion. SpaceX will reportedly pay Cursor in post-IPO stock, not cash. Cursor has been growing rapidly since 2022, reaching more than a million paying users and 4 billion in annually recurring revenue in 2026. Will the deal make xAI a top-tier AI lab? “I’m a bit skeptical on that, as xAI’s models have lagged significantly, and one strong coding model doesn’t close that gap,” writes the Pitchbook analyst Franco Granda in a new report. The deal makes sense in some ways. Cursor will get locked-in access to SpaceX’s Colossus supercomputer to train new models, which could accelerate its competitiveness relative to Anthropic’s Claude Code and OpenAI’s Codex coding tools. Cursor is an AI-native software application where programmers write, edit, run, test, and debug code. It’s trying to become a place where developers supervise fleets of coding agents rather than just work with a single agent to write code themselves. Cursor may provide SpaceX with a pathway to sell access to its models and tools to enterprises. “I believe what matters more to SpaceX is [Cursor’s] distribution of over a million daily active users . . . a distribution footprint xAI has not been able to build on its own,” Granda writes. That’s important because much of SpaceX’s pitch to investors was its plan to become primarily an enterprise AI company. ChatGPT is still the chatbot king, but rivals are gaining ground OpenAI’s ChatGPT remains the most-used AI chatbot, but Anthropic’s Claude and Google’s Gemini are catching up, according to a new report from Sensor Tower, a digital market intelligence firm. ChatGPT grew 4.8 from 1.05 billion unique monthly users in December 2025 to 1.1 billion users in May, the report says. But Gemini is gaining ground. Thanks in part to a large Android user base and deep integrations within Google’s ecosystem of products, Gemini has grown from 533 million users in December 2025 to 662 million in May 2026. (That’s a 24 uptick, for those keeping score.) Claude, meanwhile, has experienced rapid growth since the start of 2026. In December 2025 the chatbot had just 60 million monthly unique users across mobile and desktop, but that number increased fourfold to 245 million by May 2026. Sensor Tower measures unique chatbot users in 25 global markets across mobile apps, mobile web, and desktop web, so it may have the most complete picture available of chatbot market share. More AI coverage from Fast Company: Inside Fei-Fei Li’s 1 billion new AI company, World Labs The World Cup security buildout won’t end when the games do This hidden Gemini feature uses AI to teach you to be a tech savant AWS says AI agents can work on their own. It’s also building tools to keep them in line Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.