Social posts

Blog: https://maxghenis.com/blog/pulled-one-approved-another/ Prior post: https://maxghenis.com/blog/us-gutted-anthropic-rd/

X

Same stated cyber rationale, three calls in two weeks:

— Fable 5 pulled offline by a binding export-control order — GPT-5.6 cleared for ~20 companies by a voluntary request — no order — Mythos 5 reopened to 100+ organizations the government named

OpenAI’s own benchmarks put GPT-5.6 level with Mythos on cyber. Same capability, different instruments, no published rule.

maxghenis.com/blog/pulled-one-approved-another/

Bluesky

On June 12 Commerce pulled Anthropic’s Fable 5 and Mythos 5 offline worldwide. On June 26 OpenAI launched GPT-5.6 under government-approved access — kept it for its own researchers, sold it to ~20 government-approved customers, and benchmarked it by name against the two pulled models. That evening the government let Anthropic reopen Mythos 5 — to more than 100 critical-infrastructure organizations it named; Fable 5, the general-use version, stays offline.

The cleanest comparison still standing is Fable 5 vs GPT-5.6: both general-use deployments, same cyber rationale, one offline and one shipping. Three offered distinctions, all ruled out by the record:

• Capability: OpenAI’s own launch reports Sol competitive with Mythos on ExploitBench and 88.8 vs 88.0 on Terminal-Bench. The AI czar said May 8 that OpenAI’s model is “just as cyber capable as Mythos.”
• Safeguards: GPT-5.6’s system card rates its jailbreak resistance “comparable” and “directional”; the technique that hit Fable is generic and, per Katie Moussouris, “cannot meaningfully be fixed.”
• Coordination: Anthropic tested Mythos with US intelligence agencies and ran Project Glasswing before release. The lab that coordinated most still drew the mandatory order — coordination didn’t sort them either.

And the stakes compound: the cleared companies get the frontier model, their competitors don’t, and the disclosed criteria (trusted critical-infrastructure operators and agencies) don’t describe a startup.

Three models, three bespoke deals in two weeks: Anthropic under binding export-control letters (license, criminal penalties), OpenAI under a voluntary request with no order. The June 2 executive order said there would be no preclearance regime. There is one — with no published rule for who gets in.

maxghenis.com/blog/pulled-one-approved-another/

LinkedIn

Within two weeks, on the same stated concern — a frontier model’s cyber capability — the US government has pulled two AI models, cleared a third, and reopened one of the first two to more than 100 organizations.

On June 12, the Commerce Department pulled Anthropic’s Claude Fable 5 and Mythos 5 offline for every user worldwide, via an export-control order. On June 26, OpenAI launched GPT-5.6 — it kept the model running for its own researchers, shipped it to about 20 government-approved customers, and benchmarked it by name against the two models the government had just pulled. That same evening, the government told Anthropic it could redeploy Mythos 5 to more than 100 organizations that defend critical infrastructure, named by the government; Fable 5, the general-use version, stays offline.

The June 2 executive order said the government was not creating “a mandatory governmental licensing, preclearance, or permitting requirement” for AI models. It is now deciding, model by model, who may run which.

The two companies are under different authority. Anthropic’s came as signed Commerce letters from Secretary Lutnick — a June 12 export-control directive requiring a license before any “foreign person” could touch its models, with criminal and civil penalties, then a June 26 letter lifting the license for the organizations it named. OpenAI received no such order; its limit is a voluntary request, with access approved customer by customer.

The cleanest comparison is the one still standing — Fable 5 and GPT-5.6, both general-use deployments resting on the same cyber rationale, one offline and one shipping. Three explanations get offered for the difference. The public record rules out all three:

1. Capability. OpenAI made the comparison itself: GPT-5.6 Sol scores 88.8 on Terminal-Bench 2.1 (91.9 in Ultra mode) against Mythos 5’s 88.0 and Fable 5’s 84.3, and OpenAI reports Sol competitive with Mythos on the ExploitBench cyber benchmark at about a third of the output tokens. The administration’s own AI czar said on May 8 that OpenAI “now has a model that’s just as cyber capable as Mythos.” These are OpenAI’s numbers against models that were offline — but the only party claiming parity is the company the government let ship.

2. Safeguards. Fable 5 is the safeguarded deployment of Mythos. GPT-5.6’s system card describes the same categories of defense every frontier lab uses, rates its jailbreak resistance “comparable to recent predecessors,” and calls the result “directional rather than definitive.” The jailbreak that triggered the Fable order appears generic: Anthropic found no universal jailbreak and noted the same capability sits in OpenAI’s unrestricted GPT-5.5, and security researcher Katie Moussouris described the technique as something that “cannot meaningfully be fixed.”

3. Coordination. The distinction offered in reporting is that OpenAI coordinated first. Anthropic coordinated more — it tested Mythos with US intelligence agencies, ran Project Glasswing to harden critical software before release, gated Mythos itself, and proposed model-neutral disclosure rules for every frontier lab. The lab that coordinated most still drew the mandatory order. Whatever sorted these models, it wasn’t how much each company worked with the government.

The stakes aren’t abstract. The cleared companies get the frontier model; their competitors don’t. That gap can look like a few weeks, but it compounds — in my earlier post, one working day with Fable rebuilt the architecture of one of PolicyEngine’s core products. The disclosed criteria — trusted critical-infrastructure operators and agencies — don’t describe a startup, so the access gap falls hardest on the firms least able to absorb it.

Nothing on the public record separates Fable from GPT-5.6 — not capability, safeguards, or coordination — and the government has named nothing else. What decides is now in plain view, and it isn’t a rule — it’s letters: two from Secretary Lutnick to Anthropic in two weeks, and for OpenAI a request and a list of approved customers. Three models, three bespoke arrangements — with no published rule for who gets in.

Full post: https://maxghenis.com/blog/pulled-one-approved-another/