Analysis
The frontier has a guest list now. Three weeks wrote the rules.
Between June 2 and June 27, 2026, the best AI models in the world quietly stopped being things you could simply buy. An executive order, a suspension, a gated launch, and a conditional return — in under a month — established a tiered-access regime almost no one voted on. Here is what it set, and what to do if you will never be on the list.
Quick summary
- In 25 days the US assembled a working order for frontier AI: a pre-release review process, a retroactive suspension (Fable 5), a launch gated to about 20 vetted orgs (GPT-5.6), and a sector-only return (Mythos 5).
- Opinion: this is now the default, not an emergency. 'Broadly available soon,' with no date attached, is a planning assumption — not a promise.
- The only foundation no directive can revoke is open weights. The rational build: a floor you own (GLM-5.2), a stay-in-Claude option you can fail over from (Opus 4.8), and routing in between.
Twenty-five days that changed the rules
Step back from the day-to-day headlines and the shape is hard to miss. On June 2 a White House executive order told federal agencies to stand up a process for evaluating powerful new AI models before they reach the public. Ten days later that abstract process had teeth: on June 12 a Commerce Department export-control directive ordered Anthropic to suspend Fable 5 and Mythos 5 — its two most capable models — for any foreign national anywhere, including its own non-citizen staff. Unable to verify nationality in real time, Anthropic took both fully offline in about 90 minutes.
Two weeks later, June 26 did two opposite things in a single day. OpenAI launched GPT-5.6 — Sol, Terra, and Luna — but not to you: at Washington's request it went live as a limited preview to roughly 20 pre-approved organizations, through the developer API and Codex only, with individuals not eligible. The same week, the government cleared Anthropic's Mythos 5 — a different model — to redeploy, but only to a set of US organizations that run and defend critical infrastructure; it began returning by June 27. Fable 5 stayed dark for the public. (The window did close: on June 30 the US lifted the Fable 5 controls, and Anthropic restored it worldwide on July 1 — though the all-plans window ran only through July 19, and since July 20 only Max and Team Premium keep it in-plan — on Pro it's the priciest metered model Anthropic sells. The public got its model back, at a premium; GPT-5.6 is still gated. The pattern below outlived the outage.)
How frontier access got tiered, June 2026
Jun 2 — the valve
An executive order tells federal agencies to evaluate powerful new models before public release.
Jun 12 — suspended after launch
A directive pulls Fable 5 and Mythos 5 offline worldwide, three days after Fable 5 shipped.
Jun 26 — gated at birth
GPT-5.6 launches to ~20 government-approved orgs via API and Codex. Individuals not eligible.
Jun 26–27 — a conditional return
Mythos 5 is cleared for a set of US critical-infrastructure orgs. Fable 5 stays suspended for the public.
Jun 30–Jul 1 — a premium reopening
The US lifts the Fable 5 controls; Anthropic restores it worldwide — plan-included through July 7, then top-price metered credits. GPT-5.6 stays gated.
Three different gates, one direction
These were not the same mechanism, and the difference is the story. Fable 5 was pulled after launch. GPT-5.6 was gated before it — the first major American model born under the new review regime, gated at birth. Mythos came back conditionally, to a single named sector. Three different doors; one architecture behind them: the most capable models route to the government and a short list of vetted institutions first, and to everyone else later, if at all.
Our read, labeled as opinion: this is no longer an emergency response to one jailbreak. It is the start of a standing system. The June 2 order did not suspend a model — it built the valve. The three events that followed are that valve being worked three different ways in three weeks. Emergencies end. Systems compound.
Read the full analysis — free
Sign in to unlock the full Analysis column — no password. You'll also get an alert when big model changes land.
'Soon' is not a date
Both labs are saying the reasonable things. OpenAI calls the restriction a short-term step, and Sam Altman has said limits like this should not become the norm. Anthropic is working with the government and clawed Mythos partway back. Take them at their word — and still notice that none of it ships with a date you can put in a plan. A promise of 'broadly available soon' is a sentence, not a commitment.
If your product stands on a model that a policy letter can switch off in 90 minutes, or that launches to twenty companies that are not yours, your roadmap now carries a dependency you do not control and cannot schedule. That is not a reason to panic. It is a reason to design for it — before the next directive, not after.
What to actually do about it
1
Stand on weights you can hold. The one thing no directive revoked in June is the open-weight floor. GLM-5.2 shipped June 13 under an MIT license — within a few points of Opus 4.8 on coding benchmarks, at roughly a sixth of the cost — and because the weights are public, no letter can take your copy offline. Make an open model your default, not your fallback.
2
Keep a stay-in-Claude lane you can leave. Opus 4.8 was never part of the suspension and is fully available. If your work lives in Claude, keep it — but wire it so you can route around it the day it is not there.
3
Route, do not marry. The lesson of three weeks is concentration risk. The setup that survives the next directive is vendor-agnostic: a model you own the weights to, a commercial option or two behind it, and a thin layer that switches between them without a rewrite. Not sure what that looks like for your work? Ask the advisor or compare the alternatives.
Need the short answer?
Fable 5 is back worldwide as of July 1 — in-plan on Max and Team Premium, on usage credits for Pro. See the live status, or use GLM-5.2 or the new Sonnet 5 for cheaper work.
Read the brief Fabel 5 spelling guideFAQ
Can I use GPT-5.6 right now?
Almost certainly not. On June 26 OpenAI gave preview access to roughly 20 government-approved organizations, through the API and Codex only; individuals are not eligible. OpenAI says wider access is coming, but has not given a date.
Is Fable 5 back?
Yes. The US lifted its export controls on June 30 and Fable 5 returned worldwide on July 1, 2026 — and after July 19 it bills through usage credits as the priciest metered model Anthropic sells (window extended on July 8). It's back, but concentration risk (and cost) is exactly why routing still matters.
So what should I build on instead?
Don't marry one model. Keep an open-weight model you can run yourself — GLM-5.2 is the current value pick — with a Claude lane (Fable 5 for the hard jobs, Opus 4.8 or the cheaper Sonnet 5 for the rest), and route between them. See the alternatives guide or ask the advisor for a setup matched to your work.
From the Field Guide
Stop guessing which model to use
Weekly verified workflows for the credits era: which model in which seat, what it costs per task, and copy-ready prompts — plus an 8-runbook Starter Pack you can download once and keep. Every claim checked against published pricing, sources shown.
Get the free weekly briefing
One practical email a week: what's worth running on Fable 5, what it costs, and which jobs belong on a cheaper model — plus one email when the next big move lands. No spam, leave anytime.
Sources
This page is independent. Official provider pages are the source of record for access, pricing, and policy.