On Tuesday June 30, 2026, Commerce Secretary Howard Lutnick sent a letter to Anthropic informing the company that the export controls placed on Claude Fable 5 and Mythos 5 on June 12 had been lifted. Nineteen days after the harness closed, Fable 5 comes back globally today — Wednesday July 1 — on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Per Anthropic's own redeployment post, Lutnick's letter did not reissue a licence. It concluded that Anthropic no longer needs one, because a licence is a document and this deal is a set of standing obligations.

The kill switch, in other words, had a return path. Nineteen days ago that wasn't a certainty. The Loop covered the June 13 directive as an hour-and-forty-minute deployment event with no announced end date; a week later OpenAI's GPT-5.6 walked into the same clearance architecture before it had shipped a public preview at all. The story on June 13 was that the harness existed. The story on July 1 is that it can be un-set.

The letter that isn't a licence

Lutnick's statement, quoted by CoinDesk and 9to5Mac, reads: "Over the past two weeks, we have worked closely with Anthropic to analyze and approve Fable 5." Then the sentence the corporate-affairs desk is going to reread all quarter — that the company no longer needs an export license, provided it commits to a specific list of things.

Those things, per Politico via 9to5Mac, are: "proactively detect and address security risks associated with the models," work with the government on protocols for future releases, and report "malicious activity" Anthropic finds in its own systems. That is not a policy. It's not a statute. It's not even a memorandum. It is a set of continuing behavioural obligations that a cabinet secretary applied by letter and that a private company agreed to by press release. The regulatory relationship the AI policy debate has been trying to define since the summer of 2023 turned out to look, on the day it landed, like a handshake with a stationery header.

Nineteen days on the calendar

The compressed timeline, reconstructed from Anthropic's post and independent trade-press coverage:

• June 9 — Fable 5 and Mythos 5 launch. Both models share the same underlying weights; Fable 5 ships with extra safeguards for general use.
• June 12 — Amazon researchers file a report demonstrating that Fable 5 could be prompted in a way that produced disallowed output. Commerce issues an export-control directive the same day. Anthropic, unable to verify user nationality in real time, takes both models offline globally.
• June 26 — Mythos 5 clears for restoration to approximately one hundred US institutions, as part of the Project Glasswing arrangement the Loop covered on June 19.
• June 30 — Export controls lifted. Anthropic publishes the redeployment plan.
• July 1 — Fable 5 back globally on all four surfaces.

For Pro, Max, Team, and select Enterprise plans, Fable 5 usage is included through July 7 for up to 50% of the weekly quota. After that, it is usage credits. Pricing per the Let's Data Science writeup is $10 per million input tokens and $50 per million output tokens — roughly double Opus 4.8 on a per-token basis, with a 90% prompt-cache discount to soften the blow. Anthropic priced the model as a premium option and let customers self-select who actually needs it.

The 99% classifier and its false positives

The safety fix Anthropic shipped alongside the redeployment is a new classifier that, per the company's own post, "means that the specific technique described in the Amazon report is blocked in over 99% of cases." Blocked queries route automatically to Claude Opus 4.8, with a UI notification so the user can see what happened. The trade-off, Anthropic concedes, is a higher rate of false positives — coding and debugging prompts that look enough like the Amazon-reported shape get punted to the older model even when they are benign.

That is a design choice worth naming. Anthropic did not ship a hidden filter and hope nobody noticed. It shipped a visible one and said, in the same breath, that it will misfire — the Anthropic phrasing is "more benign coding and debugging requests" flagged as bypass attempts. That reads as a company writing its own bug report into the launch note. It also reads as we were told to overshoot.

The Amazon, Microsoft, Google framework

The other thing that came out of the nineteen days is a shared jailbreak-severity framework the four largest model-and-cloud shops in the US — Anthropic, Amazon, Microsoft, and Google — have agreed to draft together. Per Let's Data Science and BeInCrypto, the framework will score reported jailbreak techniques on four axes: capability gain, breadth of gain, ease of weaponisation, and discoverability.

That is the AI industry writing its own severity-scoring rubric before a regulator writes one for it. This is the same pattern the cybersecurity industry ran with the CVSS score in 2005: vendors define the vocabulary, vendors run the scorecard, vendors decide what "high" means. The good version of this outcome is a shared taxonomy that lets three labs and one hyperscaler compare notes on incoming reports without lawyering every disclosure. The bad version is that "capability gain" turns out to mean whatever the labs' own red-teams say it means, and the Commerce Department inherits the definitions along with the workflow. It will be both.

The awkward finding buried in the post

The most consequential single sentence in the redeployment post is not the 99% number. It is Anthropic's disclosure that in testing, less capable models — the post names Opus 4.8, GPT-5.5, and Kimi K2.7 — could reproduce the same findings that got Fable 5 flagged as a national-security concern in the first place.

Read that twice. The lab that just spent nineteen days offline is telling the government that the capability the export control was designed to contain was not, in fact, unique to the model that got controlled. The kill switch was aimed at the shape of a frontier model. The finding it was aimed at was reproducible on the shelf.

That is not a small caveat. It is the argument every ex-Commerce lawyer will now make on behalf of the next lab that gets flagged: what you took offline for eighteen days was a capability that the second- and third-place models had already. It is also the argument that will get attached to the industry's severity framework the moment a regulator tries to use it.

What this means

1. Standing obligations are the new export licence. The Lutnick letter dropped the licence requirement in exchange for behavioural commitments — proactive detection, coordinated release protocols, malicious-activity reporting. That regulatory shape is available for the next flagged model without another statute and without another directive. It is easier to set up and easier to reset. The workflow that took thirteen days to design around Anthropic now runs on a template.

2. The industry wrote its own rubric before the regulator did. Four vendors drafting a jailbreak-severity framework in the same fortnight the flagship model was offline is a bid to own the vocabulary of the next enforcement event. Whether Commerce adopts the four axes verbatim or reinvents them is the next fight to watch. The odds strongly favour adoption; the axes will show up in a NIST 800-series document by the fall.

3. The precedent runs both ways. Nineteen days from directive to reversal is fast enough to make the harness workable. It is also fast enough that the "we shut the lab down" threat loses force each time it fires and nothing further happens. The credibility of the mechanism is now a function of how often it gets used and against whom.

4. The frontier-vs-capable gap is smaller than the harness assumes. Anthropic's own disclosure that Opus 4.8, GPT-5.5, and Kimi K2.7 can reproduce the flagged findings is the first public admission from a lab that the "frontier model as national-security category" framing has a scoping problem. The Kimi mention matters — a Chinese open-weights release inside the same paragraph as US export-control classifications is not a rhetorical accident.

Nineteen days ago the Loop wrote that the fact pattern was being decided by the consent. It is still being decided by the consent — the consent is just now bilateral, and the second signature came with a licence-free renewal clause. The customers on the other side of the wire got what they wanted, which was the model back. The regulators got what they wanted, which was a workflow. The labs got what they wanted, which was to keep shipping. The one thing nobody in the letter got to argue is whether the harness should have existed in the first place. That argument moved to Brussels three weeks ago and hasn't been heard from since.