Effective July 8, 2026, Anthropic's updated privacy policy authorizes the collection of:
"An image of your government-issued identity document and the information appearing on it (such as your ID number and date of birth); your image in photo or video form; facial geometry templates (which may be considered 'biometric data' in some jurisdictions); and the result of the verification."
This is the restoration path for the best coding model in the world.
Eight Days
On June 9, Anthropic shipped Claude Fable 5. 95% on SWE-bench Verified. 80.3% on SWE-bench Pro. A one-million-token context window. Autonomous coding sessions measured in hours, not minutes. The most capable publicly available model ever built.
Three days later, Commerce Secretary Howard Lutnick sent a directive to CEO Dario Amodei requiring a license for the export, re-export, or domestic transfer of Fable 5 and its restricted sibling, Mythos 5. The restriction extended to any foreign person on U.S. soil. Anthropic couldn't screen users by nationality in real time. So it shut down both models for everyone on Earth.
David Sacks, co-chair of the President's Council on Science and Technology, confirmed the terms: fix the jailbreak, or de-deploy. Dario Amodei refused both. Anthropic's position — the technique exposed "minor, previously known vulnerabilities" that other models could also discover. The government overrode them.
Today is day eight. The refund deadline for subscribers who paid during the three days Fable 5 existed. No restoration date. No expiration on the directive. From the G7 sidelines, Trump said negotiations were "going fine." A proposed UK exemption collapsed. Kalshi gives 57% odds of restoration by July 1.
The first time the U.S. government has used export-control authority to pull a commercial AI model served over an API. Not chips. Not hardware. Not source code. A model accessible through a browser.
What Happened in 48 Hours
Export controls assume the restricted capability has no comparable substitutes. That assumption broke in 48 hours.
Within one week, four open-weight coding models shipped. Cohere pushed North Mini Code. Moonshot released Kimi K2.7-Code the same day as the ban. Zhipu shipped GLM-5.2 — 744 billion parameters, MIT-licensed, one-million-token context, within one point of Claude Opus 4.8 on long-horizon coding tasks, at a tenth of the cost. Poolside released Laguna M.1 under Apache 2.0 — 72.5% on SWE-bench Verified.
Meanwhile, OpenRouter Fusion demonstrated that combining cheap models through a judge-synthesizer pipeline scores 69% on the DRACO research benchmark. Fable 5 solo: 65.3%. A budget panel — Gemini 3 Flash, Kimi K2.6, DeepSeek V4 Pro — scored 64.7%. Within one point of Fable, at half the cost.
Fable 5 still holds the #1 position on DeepSWE — 70% PASS@1, three points ahead of GPT-5.5. The best autonomous coding model in the world is the one that's offline. But the ecosystem didn't wait for it to come back.
The Escalation
Here is what accessing frontier AI has cost over the past two years:
The verification is processed through Persona Identities, a third-party KYC platform — the same infrastructure banks use for anti-money-laundering compliance. Accepted documents: passports, driver's licenses, national identity cards. Photocopies and digital IDs are rejected. Age verification runs through Yoti, using facial age estimation or document scans.
And it applies only to consumer users — Claude Free, Pro, and Max. Enterprise and Platform customers operating under Commercial Terms are exempt. The individual developer — the person who benefited most from Fable-class autonomous coding — is the one who submits their face. The enterprise with a legal department doesn't.
What This Is
This isn't Anthropic's fault. They didn't choose this. The biometric verification infrastructure is their attempt to restore access, not restrict it — to give the Commerce Department what it needs to allow domestic-only availability. The alternative is Fable 5 staying offline indefinitely.
But intention doesn't change what the requirement is. For two years, the cost of frontier AI was measured in dollars per million tokens. Then in enterprise contracts and compliance frameworks. Now in facial geometry templates.
The company preparing for a $965 billion IPO is simultaneously building KYC infrastructure so that American developers can prove they're American before accessing a coding tool. No frontier AI subscription agreement was designed for government-ordered suspension. That gap won't be fixed by one refund cycle.
GLM-5.2 is MIT-licensed. Laguna M.1 is Apache 2.0. Neither requires your passport.