The Permission Tax
is here

The U.S. government just got a 30-day head start on the most powerful AI models on Earth. On June 26, 2026, OpenAI began a limited preview of GPT-5.6 Sol, Terra, and Luna. Then it locked the door. Only a small group of government-approved partners can use them right now.

This is not a leak. OpenAI announced it openly. The company said it limited access "at the request of the U.S. government." Two weeks earlier, Anthropic got a directive at 5:21pm Eastern on June 12 and shut down two of its own models, Fable 5 and Mythos 5, by that evening.

Here is the part that should worry every builder. Your skill no longer decides if you can build with the best models. A government review does. The most capable tools on the planet are now behind a velvet rope, and the bouncer works for a federal agency.

The Permission Tax

I call this shift the Permission Tax. It is the new cost every frontier-model builder pays, and you do not pay it in dollars. You pay it in clearance.

For most of software history, access scaled with ability. You learned to code, you got an API key, you shipped. The barrier was your brain, not your paperwork.

That model is now splitting in two. Tier 1 gets the strongest models first: federal agencies, critical infrastructure operators, and hand-picked "trusted partners." Tier 2 is everyone else, waiting weeks for the broader release, or shut out entirely.

The trigger is Executive Order 14409, signed June 2, 2026. It directs agencies to build a classified benchmarking process within 60 days to label some models "covered frontier models" based on cyber capability. Once labeled, the developer can give the government up to 30 days of access before anyone else. Then the government helps pick who gets in next.

Here is the trap. The order swears it creates no licensing or preclearance regime. The White House text says nothing "shall be construed to authorize" a mandatory permit. But as Lawfare put it, "voluntary" looks different when your customer is the federal government. When the buyer holds the gate, the gate is not really optional.

The Allowlist Becomes the Moat

Pull this apart slowly, because the structure matters more than any single launch.

Three forces are stacking on top of each other. First, OpenAI's own capability tiers: Sol is the strongest, with Terra and Luna below it. Second, the classified cyber benchmark to be built by the NSA, CISA, and Treasury, which decides what counts as a "covered frontier model." Third, the human approval layer, where access gets cleared customer by customer.

That last layer is the quiet revolution. A startup founder with a brilliant product can be perfectly capable of using GPT-5.6 Sol. But if their customer profile or jurisdiction fails a government-mediated review, they sit in Tier 2. Capability stops being the currency. Legibility becomes the currency.

Think about who passes a vetting process fastest. Big firms with compliance teams. Established players already wired into national-security work. The companies that are easy to read and easy to trust. This is the asymmetry that should keep founders up at night.

Early access is becoming a moat, and not the kind you build. It is the kind the state hands you. Anthropic's case shows the harder edge. Access went from a product decision to a negotiation with officials.

Here is the contrast that defines the moment. Amateurs ask, "Can I build it?" Operators in this new regime must ask, "Am I cleared to build it?" Those are different questions, and only one of them used to matter.

I think the deepest risk is not the 30-day delay. Delays end. The deeper risk is dependency. If you build your company on a frontier API, you now inherit changing prices, changing safety filters, and changing access set by government pressure. Anthropic's Mythos 5 got pulled after a reported jailbreak allegedly unlocked unintended capabilities. Defenders who relied on it lost their tool overnight.

The data is mixed on whether this stays narrow. The gating could be temporary: preview for the vetted few, then broad release for all. Or it could harden into a permanent class system. The precedent is being set in real time, case by case, not by clear statute. That is exactly how soft rules become hard ones.

There is a real safety argument on the other side, and I take it seriously. The NTIA has warned that open-weight foundation models could plausibly lower the barrier for CBRN weapons and enable offensive cyber operations. If a model can find and exploit vulnerabilities at scale, treating its release like normal software is naive. Some controlled rollout is defensible. The question is whether discretion without published criteria or appeals is the right tool.

2031

Three signals inside the same shift

Zoom out five years. The compounding effect of the Permission Tax is what decides the next decade, not the June 2026 headlines.

Consider Google DeepMind's reaction. On June 25, 2026, it reorganized its AI Coding Strike Team and expanded its scope to a midtraining phase, reportedly following planned departures of key staff. That is what acceleration under pressure looks like. The labs are restructuring internally because the competitive ground is moving fast.

Now layer the gate on top. If the strongest models flow first to a regulator-aligned tier, market share clusters there too. High-margin cybersecurity work, critical-infrastructure deployments, defense-adjacent tooling. Small in headcount, huge in value per customer. The frontier market splits by 2031 into a permissioned segment and a mass developer segment.

This is counterpositioning in reverse. Usually the startup uses a tactic the incumbent cannot copy. Here the incumbent and the state share an advantage the startup cannot replicate: political legibility. That is an asymmetric disadvantage for new entrants, and it compounds.

So where is the durable edge? It is not in chasing frontier access you may never get. It is in building on what stays open. Open-weight models, prior-generation models, model-agnostic architectures that route across vendors. Only cash is real, and only access you control is durable. The rest is somebody else's permission slip.

It is unclear whether the open ecosystem stays competitive enough to matter. But the bet that survives every regime change is the one that does not depend on a gate you do not own.

What to Build This Weekend

Stop waiting for a frontier key you might never get. Build something that runs on models you can actually access today. Get your reps in on the open stack.

First, pick one open-weight model you can self-host or call freely. Something like a strong open model you can fine-tune. Fine-tuning just means feeding it your own examples so it specializes in your task. Run it locally or on a cheap GPU host.

Second, build a model-agnostic router. This is a thin layer of code that sends your request to whatever model is available, then swaps providers if one disappears. Write it so changing models takes one line, not a rewrite. This is your insurance against the next 5:21pm directive.

Third, test the failure case on purpose. Disable your primary model and confirm your app falls back to a backup without breaking. Things break in this regime. Build like they will.

You do not need a CS degree or a clearance for any of this. You need a weekend and a willingness to learn in public. The builders who win the next five years are the ones who never let one gatekeeper hold their entire stack. Build one tiny thing this weekend that survives without permission.