Week of 2026-06-22 to 2026-06-28 · The US didn’t switch a model off this week. It decided who’s allowed to switch one on — and OpenAI said yes.

Two weeks ago the government switched off a model. This week it did something more durable: it decided who gets to turn one on.

On June 26, OpenAI previewed GPT-5.6 Sol, its most capable model. It did not ship to a billion ChatGPT users. It shipped to roughly 20 partners whose names the US government approved one at a time. Bloomberg reported the administration asked OpenAI to stagger the release and signed off on access customer by customer. It is the first time an American company has launched a frontier model under a government-managed access list.

That is the story of the week. Not the model. The list.

”Voluntary” is doing a lot of work

The mechanism is Trump’s June 2 executive order, “Promoting Advanced Artificial Intelligence Innovation and Security.” On paper it is light-touch. Developers may voluntarily submit a “covered frontier model” for up to 30 days of pre-release review, with the NSA director setting the cyber-capability threshold. The order explicitly disclaims any “mandatory governmental licensing, preclearance, or permitting requirement.” No license. Just a friendly look.

GPT-5.6 Sol was the first real test of that framework, and “voluntary” did not survive contact. Per The Information, the implementation required per-customer approval during preview — Altman reportedly told staff the administration would approve customers individually. So a company with about a billion users shipped its best model to a 20-name list the government vetted. There is no licensing regime. There is just a release nobody could make without sign-off.

Watch the two-step. On June 2 Altman posted that “the new EO gets the balance right.” On June 26 OpenAI wrote that it does not believe “this kind of government access process should become the long-term default.” Endorse the frame, then object to the frame once you’re inside it. That’s the shape of a soft regime tightening: the disclaimer says “not mandatory,” and the practice says “you’ll wait for the list.”

The gate catches its own builders

The same week confirmed who the gate is willing to catch: everyone, including the people who built it.

Parts of the NSA lost access to Anthropic’s Mythos 5 — collateral from the June 12 export order that barred foreign nationals from Fable 5 and Mythos 5 and forced Anthropic to pull both globally. Analysts were notified on a Friday that the model was going dark. The agency that the EO names as the authority on which models are too dangerous to release lost access to one of them through a supply-chain dispute.

The justification feeding the gate is itself contested. Senator Mark Warner said NSA Director Joshua Rudd told him Mythos “broke into almost all of our classified systems, not in weeks, but in hours.” A US official then told Warner he’d misunderstood — that was a sanctioned red-team test of internal networks, not an escape. So the headline capability (“AI breached the NSA”) that makes a guest list feel reasonable is, on the record, a misread of a test the NSA ran on itself. The fear is doing the gating; the fact is still in dispute.

This is the switch-off thread, escalated

We have tracked this for a month. In W24 safety became the weapon: the danger narrative Anthropic authored got turned against it, and Commerce switched off its two best models. In W25 we watched the demand route around the ban to open weights. Underneath both sits the export-control dive: you cannot control a trained model, because the weights are just numbers and the numbers leave.

W26 is the next move, and it is qualitatively different. W24 was negative control — taking one lab’s model away, against its will. W26 is positive control — deciding who may receive a model, with the lab’s cooperation. The state went from confiscating Anthropic’s distribution to managing OpenAI’s. And OpenAI volunteered. That is the escalation worth marking: the frontier labs are starting to accept government-managed release as the price of shipping, and they’re calling it voluntary while they do it.

The wall has no back

Here is why the guest list is incoherent, not just heavy-handed: the capability it gates is, at the same time, becoming a free download.

The day after Sol’s gated preview, Asian startups shipped models built explicitly to fill the Mythos gap. Tokyo’s Sakana AI launched Fugu, which it claims “stands shoulder-to-shoulder with leading models like Anthropic’s Fable 5 and Mythos Preview.” China’s 360 unveiled Tulongfeng for vulnerability discovery and Yitianzhen for cyber defense. 360 founder Zhou Hongyi named the dynamic precisely: the danger is “one-way transparency,” where some actors get advanced vulnerability-finding capability and others are locked out. That is the exact asymmetry a US access list manufactures. It doesn’t deny the capability to adversaries. It denies it to the people on the wrong side of the list — who are mostly Americans and allies — while the capability ships anyway from Tokyo and Beijing.

Add last week’s MIT-licensed GLM-5.2, now level with GPT-5.5 on real-world agentic work, and DeepSeek’s permanent price floor from June 28. A government access list works exactly as long as the gated model is meaningfully better than the ungated substitute. That gap is the whole ballgame, and it is closing. Build a beautiful front door with a 20-name guest list, and it does nothing if the back wall is missing.

What this means if you build on these APIs

Last week the lesson was: your best model can be switched off by someone who is neither your vendor nor your government. This week sharpens it: your access to the frontier is now downstream of a guest list you are not on.

The concrete change is access latency, not a hard wall. GPT-5.6 Sol will reach general availability “in the coming weeks” — but the most capable tier shipped first to approved partners, and that staging is now a thing the government can ask for on any frontier launch it deems “covered.” Plan for it. Assume the next flagship you depend on may arrive gated or delayed for reasons that have nothing to do with the vendor’s readiness.

The hedge is the same one this column has argued for all quarter, now with a second reason. Keep an open-weight fallback wired up and continuously eval’d — not because it wins, but because nobody vets you to use it. Provider-portability priced in switch-off risk. Now it prices in access-list risk too. The open floor — GLM, Fugu, DeepSeek — is not where serious work goes. It’s the exit you keep unlocked because the front door now has a bouncer.

Also this week

  • Anthropic emptied a wing of DeepMind. Nobel laureate John Jumper (AlphaFold lead) and Gemini researchers Jonas Adler and Alexander Pritzel went to Anthropic; Arthur Conmy followed; Noam Shazeer went to OpenAI. Alphabet shed roughly $270B over two sessions. The frontier is concentrating people, not just compute — a private mirror of the public gating above.
  • OpenAI became a chip company. It unveiled Jalapeño, a custom inference ASIC built with Broadcom, designed in nine months, deploying at gigawatt scale by end of 2026 with Microsoft taking 40% of first production. This is the week’s other structural move and gets the deep dive — see below.
  • The power bill found the ballot box. Utah Senate President J. Stuart Adams lost his primary on June 25 after backing a data-center project; a fellow commissioner said the vote “cost me the election.” Polling puts local opposition near 70%. The “who pays for AI’s power” thread stopped being a forecast.
  • Claude joined the org chart. Anthropic shipped Claude Tag, an always-on Slack teammate with a single shared identity across a company. An agent that is a named coworker is a different consent and audit problem than a tool you invoke — worth watching.
  • The labor data kept darkening. 2026 tech layoffs neared 150K at ~900/day, with entry-level dev employment (ages 22–25) down nearly 20% since 2024 — profitable firms cutting headcount to fund AI capex. The “affordability crisis” thread on HN (254 comments) is the same question from the demand side: who’s actually paying for all this.
  • Claude Code shipped a real undo. v2.1.191’s /rewind resumes a session from before /clear ran — directly answering the top context-management pain we covered in the context-budget dive. Aggressive clearing is now recoverable.

Ledger note: nothing came due this week. The W23 Copilot-walkback call is due ~Jul 5 (still open, no reversal seen). The W24 export-ban-narrowing call is due ~Aug 14 — and this week cut against it: the NSA losing Mythos and Asian clones filling the gap are evidence the ban is dragging, not narrowing. Scorecard stays 0–0.

One thing to watch

GPT-5.6 Sol’s guest list is either an incident or a regime. I think it’s a regime.

Prediction (60% confident): by 2027-Q1, at least one more flagship launch from a major US lab (OpenAI, Anthropic, Google, or xAI) ships under a government-staggered or government-approved access arrangement before general availability — a second instance, confirming the June 2 EO’s “voluntary” review has hardened into a release gate rather than a one-off. If the next round of frontier launches goes out to the public on day one with no federal staging, I’m wrong, and Sol was the exception.

Deep dive this week: OpenAI’s Jalapeño is the most concrete sign yet that the AI labs are becoming chip companies. The piece asks why — and what “owning the full stack” actually buys when the rail underneath the model is the last layer Nvidia still owns.