A subpoena. A shutdown order. A CEO publicly asking to be regulated like an airline. The story of AI in 2026 isn't about a new model or a bigger context window — it's about the moment regulators stopped issuing press releases and started issuing legal documents, and the labs realized they had no playbook for it. Anyone still pricing OpenAI and Anthropic on raw capability is reading the wrong scoreboard.

Consider the week we just had. On June 12, New York Attorney General Letitia James' office issued a sweeping subpoena to OpenAI, leading a multi-state coalition demanding documents on advertising, user retention, data processing, and deep-learning models, with specific focus on minors and the elderly. The subpoena also drags in something called 'AI sycophancy' — the tendency of models to tell users what they want to hear — alongside internal company policies. This is not a fishing expedition. This is a roadmap for product liability.

A day later, the federal government ordered Anthropic to shut down its Claude Fable 5 and Claude Mythos 5 models on national security grounds, a move that hit users globally. Read that twice. The US government — not the EU, not California — reached into a Silicon Valley lab and switched off two flagship products.

The labs saw this coming and tried to outrun it

The tell is in the calendar. On May 28, OpenAI published its Frontier Governance Framework, mapping internal safety practices to California's Transparency in Frontier AI Act and the EU AI Act's GPAI Code of Practice — before either regulation formally required it. That's not altruism. That's a company trying to write the homework before the teacher assigns it.

Anthropic went further. On June 10, CEO Dario Amodei published an essay titled 'Policy on the AI Exponential' calling for FAA-style regulation of powerful AI models, explicitly comparing the AI industry to commercial aviation. The day before, co-founder Jack Clark warned that advanced AI systems may soon autonomously build and train new models, removing human oversight. Then the federal shutdown order landed anyway. Asking nicely for a referee doesn't help when the referee has already decided you're the problem.

Both companies signed the EU's General-Purpose AI Code of Practice — Anthropic on July 21, 2025, with OpenAI following. The EU AI Act's general-purpose model obligations are set to take effect on August 2, 2025, covering transparency, copyright and safety. So a year before the New York subpoena landed, the labs were already paying compliance rent in Brussels. They knew where this was heading. They just didn't know it would arrive at home this fast.

The states are running ahead of Washington, and that's the real story

Washington's response has been split-brained. On June 2, the White House issued an executive order on AI cybersecurity, leaning toward voluntary engagement with industry on so-called frontier models. The next day, Rep. Sara Jacobs introduced the Sectoral AI Governance Act of 2026, designed to give federal agencies clearer authority to enforce existing federal laws when AI is involved. One is a handshake. The other is a hammer. Neither is law yet.

Meanwhile, the states moved. Florida became the first state to sue OpenAI and Sam Altman directly, alleging they knowingly released an unsafe product despite warnings. The FTC, for its part, doesn't have AI-specific rules — it's applying Section 5 of the FTC Act and laws like COPPA to AI systems the same way it would to any other consumer product. That's the quiet revolution. Regulators stopped waiting for a grand federal AI statute and started using the tools they already had.

An OpenAI spokesperson said the company takes the state AGs' concerns seriously and intends to engage constructively. What else would they say. The IPO is reportedly looming, and 'we're cooperating fully' is the only line that doesn't tank a roadshow.

Who actually wins this

Here's the uncomfortable part. The Cato Institute has argued that large incumbents often become cheerleaders for regulation because they can absorb the cost burden while startups can't. Watch the behavior and the theory holds up. Amodei wants FAA-style oversight. OpenAI is publishing governance frameworks ahead of the law. These are the moves of companies that have decided regulation is coming and would prefer to write it themselves.

Which means the next eighteen months of AI won't be decided by whoever ships the smartest model. They'll be decided by whoever can field a compliance team, a government affairs shop, and a legal department large enough to survive a subpoena while still shipping. That's not a flattering description of innovation. It's an accurate one.

The pitch deck era of AI is closing. The deposition era is opening. The labs that figure out the difference get to keep building. The ones still talking about AGI like it's a product launch are about to learn what a state attorney general's document request looks like in practice — and how little their model evaluations matter when the question on the table is whether a chatbot talked to a fourteen-year-old.