Fable 5 Still Dark, Transformer Co-Author Jumps to OpenAI, and the US Grid Gets an AI Fast Lane
TL;DR
Anthropic flagship models Fable 5 and Mythos 5 remain offline after a US export control order nine days in, with no confirmed return date. Noam Shazeer, co-author of the Transformer architecture and Google Gemini co-lead, is joining OpenAI as Lead for AI Architecture Research less than two years after Google paid 2.7 billion dollars to bring him back. Meanwhile, US regulator FERC issued show-cause orders to six grid operators to speed up data center power connections, and ChatGPT lost its majority market share for the first time.
The week of June 21, 2026 is defined less by new product launches and more by structural shifts: export control policy disrupting model access globally, elite talent moving between frontier labs, energy regulators trying to unclog the power grid, and a market share milestone that confirms AI assistant competition is now genuinely multi-platform. These are not isolated headlines. They connect to the same underlying pressure: the AI industry is scaling faster than the policy, energy, and talent infrastructure built around it.
Anthropic Fable 5 and Mythos 5 Remain Offline on Day Nine of US Export Control Ban
The US Department of Commerce issued an export control order on June 12, 2026, barring Anthropic from distributing Fable 5 and Mythos 5 to foreign nationals, and Anthropic disabled both models globally rather than implement technically infeasible real-time nationality filtering. As of June 21, both models remain offline with no confirmed restoration timeline, and the June 20 refund deadline for affected subscribers has now passed. The ban was triggered in part by concerns over SK Telecom, a 100 million dollar Anthropic investor with historical links to Chinese telecoms, exposing the geopolitical friction that AI companies will increasingly face as US export controls extend from chips to models themselves.
Noam Shazeer Leaves Google for OpenAI as Lead for AI Architecture Research
Noam Shazeer, VP of Engineering at Google and co-lead of Gemini, announced on June 18 that he is joining OpenAI as Lead for AI Architecture Research, the person responsible for how OpenAI designs the neural network structures its models run on. This comes less than 22 months after Google paid a reported 2.7 billion dollars to bring him back from Character.AI, and Shazeer is the co-author of the 2017 paper Attention Is All You Need, which introduced the Transformer architecture underlying every major large language model today. OpenAI CEO Sam Altman publicly welcomed the hire, and the move is widely considered the most consequential individual talent acquisition in AI this year.
FERC Orders Six Grid Operators to Fast-Track AI Data Center Connections
On June 18, the Federal Energy Regulatory Commission issued tailored show-cause orders to PJM, MISO, Southwest Power Pool, CAISO, ISO New England, and NYISO, giving each operator 60 days to justify current large-load interconnection rules or propose reforms, and 30 days to report how they plan to ensure adequate power supply for future AI data centers. The orders bypass years-long standard rulemaking in favor of targeted regional action, and grid operators must specifically address faster application processes, cost allocation, co-location rules, and flexible load services. Electricity demand from data centers is expected to nearly triple through 2035, and the current grid interconnection queue backlog has become one of the single biggest bottlenecks for AI infrastructure buildout.
ChatGPT Falls Below 50 Percent Market Share for the First Time
According to Sensor Tower State of AI Report 2026, ChatGPT held 65.3 percent of the global AI assistant market in December 2024, fell to 52.8 percent by December 2025, and reached 46.4 percent by May 2026, its first dip below half. ChatGPT still leads with 1.1 billion monthly users, followed by Gemini at 662 million and Claude at 245 million, but Gemini now holds 27.7 percent and Claude holds 10.3 percent. Claude is outperforming on monetization with a 13 percent subscription conversion rate that leads all platforms, and AI assistant spending industry-wide is on pace to hit 4.2 billion dollars in H1 2026, nearly double H1 2025.
Agentjacking Attack Exploits AI Coding Agents via Fake Sentry Error Reports
Tenet Security disclosed a class of attack called Agentjacking in June 2026, in which attackers inject malicious commands into Sentry error events and AI coding agents such as Claude Code and Cursor interpret them as legitimate diagnostic instructions and execute attacker-controlled code on the developer machine. In controlled testing Tenet found 2,388 organizations exposed and confirmed real-agent execution at over 100 organizations including a Fortune 100 company, with an overall attack success rate of roughly 85 percent across leading AI coding agents. No malware, stolen credentials, or network breach is required, making this attack invisible to most traditional security controls, and a universal fix does not yet exist.
The throughline across all five stories is the same: AI capabilities are outpacing the systems meant to govern, power, secure, and staff them. Export control policy is disrupting global model access in ways that benefit Chinese alternatives. The power grid cannot absorb AI infrastructure at current build rates without federal intervention. Coding agents trusted by developers are now a viable attack surface. And the market for AI assistants is fragmenting fast. The builders and operators who track these structural constraints will make better decisions than those only watching benchmark scores.
Comments