The Penalty for Being Good
How the US government is accidentally handing China the AI race, and why the most principled company in AI is paying the price
On June 9, 2026, Anthropic released Fable 5. By nearly every benchmark, it was the most capable AI model ever made available to the general public. Three days later, the US government ordered it offline.
The official reason was a jailbreak. On the evening of June 11, Amazon CEO Andy Jassy personally called Treasury Secretary Scott Bessent and other senior administration officials to relay findings from Amazon’s security researchers. Those researchers had used a series of prompts to get Fable 5 to provide information useful for cyberattacks. Luta Security CEO Katie Moussouris, who reviewed the underlying research paper, described the technique as asking the model to read a specific codebase and identify software flaws, work she said is “exactly what the model was intended to do” for defensive purposes. By the following afternoon, Commerce Secretary Howard Lutnick had issued a letter to Dario Amodei invoking national security export controls. Anthropic had 90 minutes to comply. The company received no prior notification of any national security concern.
The government called it a security emergency. The evidence suggests something more complicated.
The feedback loop nobody is talking about
Pulling a frontier AI model from hundreds of millions of users is not just a business disruption. It shuts off the most powerful feedback mechanism in AI development. Real users surface failure modes, edge cases, and capability gaps that no controlled testing environment can replicate. That signal feeds directly into the next training run. When the model goes offline, so does the engine that makes the next version better.
China does not have this constraint. DeepSeek, Alibaba, ByteDance are releasing, iterating, and accumulating signal at full speed. Every month Anthropic spends offline is a month Beijing’s labs spend learning from real usage at scale.
The gap is already narrower than most people realize. At the end of 2023, the best closed model scored around 88% on MMLU while the best open alternative managed roughly 70.5%, a gap of 17.5 percentage points. By early 2026, Epoch AI found that open-weight models trail the state-of-the-art by roughly three months on average, down from nearly a year in late 2024. Multiple benchmark families have converged to single-digit gaps, and on coding tasks, open-weight models from Chinese labs now match or beat several closed frontier models.
The open-weight ecosystem is not chasing the frontier anymore. It is standing on its doorstep.
Every regulatory delay on a frontier lab is a gift to that ecosystem, and a significant portion of it is Chinese.
The government that fell behind
The natural assumption, when the government pulls a commercial AI model citing national security, is that something better exists in reserve. Some classified system the public cannot see.
There is not.
In November 2024, Anthropic became the first AI company to clear the Pentagon’s classified hurdles, deploying Claude through Palantir’s systems at the “Secret” classification level. No other frontier model was cleared for classified use. When Fable 5 launched, the most advanced AI on the planet was a commercial product, and the entity responsible for evaluating its risk was relying on briefings from a competitor’s security team to understand what it was dealing with.
For decades, the US government ran technology programs years ahead of the public. The Manhattan Project, the SR-71, stealth aircraft. AI has inverted every condition that made that possible. The frontier is now funded by private capital. Research is published openly because talent acquisition depends on it. The training data is the public internet. There is no chokepoint the government controls at the level of capability.
What the government can control is who gets to release. That is the lever it pulled, against a company it had been at war with for months, in a direction that consistently benefited one competitor over another.
A timeline that tells a story
February 24, 2026. Defense Secretary Pete Hegseth delivers a formal demand to Dario Amodei: remove all usage restrictions and grant the Pentagon access to Claude “for all lawful purposes,” with no exceptions. The consequences for refusal are explicit: termination of a $200 million contract, supply chain risk designation, and potential invocation of the Defense Production Act. The demand arrives one day before the US and Israel launch a large-scale offensive against Iran.
February 27, 2026. The deadline expires. Anthropic holds its two red lines: no mass surveillance of American citizens, no fully autonomous weapons without human oversight. President Trump orders every federal agency to immediately cease using Anthropic’s technology. Hegseth designates Anthropic a “supply chain risk,” a classification historically reserved for foreign adversaries like Huawei. It is the first time that designation has been applied to an American company.
February 27, 2026, hours later. OpenAI announces a Pentagon deal. CEO Sam Altman acknowledges the negotiations were “definitely rushed.” OpenAI employees who had signed an open letter supporting Anthropic’s position revolt publicly. The company’s top robotics executive resigns. Claude surges past ChatGPT to become the most-downloaded app in the Apple App Store.
Within hours of Anthropic being blacklisted, its primary commercial rival had the contract.
March 2026. Anthropic files suit in both the Northern District of California and the D.C. Circuit Court of Appeals, challenging the supply chain designation as unlawful and retaliatory. A federal judge blocks the designation, calling the government’s actions “arbitrary” and “capricious.”
April 2026. The White House quietly begins working to walk the designation back. Senior officials meet with Amodei. One source describes the effort as a way to “save face and bring em back in.” The NSA, never covered by the ban’s restrictions, is using Anthropic’s Mythos model. The government needs what it tried to destroy.
June 9, 2026. Fable 5 launches. Three days later it is gone.
On June 13, one day after the export directive, Hegseth posts on X: “Three months ago, @DeptofWar kicked @AnthropicAI out of our building, forever. Every passing day proves why that was the right move.”
What Anthropic refused, and what happened next
Anthropic’s two red lines were narrow. No mass surveillance of American citizens. No fully autonomous weapons without a human in the decision chain.
The day after Anthropic held those lines, the US and Israel launched Operation Epic Fury against Iran. Within 48 hours, a US Tomahawk missile struck the Shajarah-Tayyebeh elementary school in Minab, Iran. Rescuers pulled 175 bodies from the rubble. Preliminary reports indicated the school was misidentified as a military site, potentially through AI-assisted targeting analysis processed through Palantir, the same defense contractor through which Claude had been deployed on classified networks since November 2024.
The Pentagon, which had publicly banned Anthropic, was still using Claude. Military commanders told reporters they would continue regardless of the president’s order until a viable replacement emerged. “Whether his morals are right or wrong or whatever,” one source told the Washington Post, “we’re not going to let Dario Amodei’s decision-making cost a single American life.”
Amodei said: “Disagreeing with the government is as American as apple pie.”
OpenAI’s contract, signed the same afternoon Anthropic was blacklisted, allows its models to be used for “any lawful purpose.” Anthropic had sought explicit contractual bans on surveillance and autonomous weapons regardless of what the government argued was legal. OpenAI deferred to existing law and policy, the same framework that allowed surveillance practices revealed by Edward Snowden to operate for years before courts ruled them unlawful.
One of OpenAI’s own safety researchers publicly called the approach “window dressing.” MIT Technology Review put it plainly: “Anthropic pushed for moral boundaries. OpenAI settled for softer legal ones.”
The political economy of a takedown
Since early 2025, the Trump administration has been in active negotiation with OpenAI about a government equity stake, a deal that would give the US a financial interest in OpenAI’s success ahead of an IPO. Altman first pitched the concept directly to Trump in early 2025. Talks have continued for more than a year.
Anthropic is explicitly excluded from those conversations. A source confirmed the company is not having equity discussions with the administration, not by choice but as a consequence of being frozen out entirely.
OpenAI is also linked to a PAC called “Leading the Future” that has spent heavily in support of the current administration. Anthropic has no equivalent political infrastructure.
The research behind the Fable 5 directive came from Amazon, Anthropic’s largest investor at roughly $13 billion committed, and a direct cloud competitor with its own AI ambitions. Jassy personally carried those findings to the Treasury Secretary. Commerce Secretary Howard Lutnick, who has a documented adversarial relationship with Anthropic, issued the directive. Hegseth, who designated Anthropic a supply chain risk in February, publicly celebrated the Fable ban the following day.
None of that constitutes proof of a coordinated effort. But it is a set of documented facts in which Anthropic is consistently the target of punitive action while its primary competitor consistently benefits, with no transparent process that would allow the public to evaluate why.
The company that refused to let its AI help identify bombing targets is the one the government called a supply chain risk. The company that signed within hours of the blacklisting got the contract and the equity conversation.
The trap
Anthropic’s safety-first positioning, the thing that made it the most trusted AI company in the world to many of its customers, is also what made it the most visible regulatory target.
It publishes what it finds. It draws lines and holds them publicly. A lab that does not talk about catastrophic risk does not get government directives about catastrophic risk. The more seriously Anthropic took safety, the larger a target it painted on itself.
Meanwhile, the open-weight ecosystem faces none of these constraints. Open-weight models, once released, cannot be recalled by any government directive. They have no terms of service, no export controls, no CEO who can be summoned to the Pentagon. Within the same week Fable 5 was pulled, Chinese open-weight models stepped into the gap. Enterprises redirected API calls without rebuilding their infrastructure, because model endpoints are now interchangeable enough that swapping one out is a configuration change.
The government’s action, intended to contain a capability, accelerated adoption of that capability’s successors, built by the people the policy was ostensibly protecting against. Cold War export control logic applied to technology the government does not lead cannot produce a different result. You can only embargo a chokepoint you control.
What this is actually about
Anthropic drew two lines and held them while a war was starting, under explicit threats of contract termination and foreign adversary designation. The company held them because it believed, and still believes, that frontier AI operating without human oversight in lethal contexts is not something the technology can yet safely do.
That position may be wrong. Reasonable people disagree on where exactly the line should sit. But it is a position held in good faith, stated publicly, and backed by safety research that no other company in the world has done at comparable depth.
The punishment was a supply chain risk designation historically reserved for Huawei. The reward for abandoning those same positions was a classified Pentagon contract and an equity negotiation with the White House.
China is not slowing down. Their labs are not being pulled three days after launch. Their government is not designating their best AI companies foreign adversaries. And their open-weight models have closed to within a few months of the frontier.
Anthropic drew the lines. The government erased them, handed the contract to a competitor, and called it national security.
Sources: CNN, NPR, Fortune, MIT Technology Review, The Conversation, Democracy Now, OSV News, Times of Israel, CNBC, NBC News, Axios, Scientific American, Semafor, Epoch AI, ExplainX, Ability.ai, Built In, InfoQ, MLQ.