Generated by Codex with GPT-5

This summary covers The Economist’s June 20th, 2026 Briefing article listed in the contents under Anthropic and published under the headline A test of faith.

The article presents Anthropic not just as another artificial-intelligence startup, but as a company whose moral self-image has become part of its power. Founded by defectors from OpenAI who believed frontier AI needed stricter safeguards, Anthropic built its identity around safety, caution and a quasi-missionary belief that artificial general intelligence could reshape civilization. That stance once made it look like the pious sibling in the AI race. Now that its models are commercially valuable and strategically sensitive, the same posture has put it in direct conflict with the American state.

The immediate dispute is over Fable 5 and Mythos 5, Anthropic’s newest and most capable models. The Trump administration ordered the firm to prevent non-Americans from using them, including foreign employees inside Anthropic itself, citing national-security concerns. The order followed claims that Fable could be jailbroken and that Mythos had alarming cyber capabilities. In practice, Anthropic could not separate access neatly by nationality, so it suspended use more broadly. A fight over safety controls quickly became a larger fight over who gets to decide how the world’s most powerful AI is deployed.

A Safety Lab Becomes A Power Broker

Anthropic’s founding story matters because it explains why the company is unusually exposed to political conflict. Dario Amodei and his co-founders left OpenAI in 2021 because they thought the leading lab was not serious enough about safety. Anthropic’s employees talk about ethics and long-term risk more openly than most technology firms, and Amodei has repeatedly called for binding AI regulation. The article compares this posture to religious authority: powerful in moral terms, but vulnerable when secular rulers decide to intervene.

That moral authority has not slowed the business. If anything, it helped Anthropic find a profitable niche. Rather than chasing mass-market chatbot fame, the company concentrated on business customers, especially software developers. Claude Code, its coding agent, became popular because it can handle tasks that would take humans hours or days and because its output can be checked. Corporate customers also liked Anthropic’s emphasis on safeguards, which made adoption feel less reckless.

The result is that Anthropic has become one of the most valuable AI firms in the world. The article says both Anthropic and OpenAI have filed for IPOs and could be valued at around \$1trn. Anthropic’s annual recurring revenue has surged, and its business customers account for a much larger share of income than OpenAI’s. That gives it a clearer path to profits, but it also raises the stakes. A company that once sounded like a safety watchdog is becoming a potential gatekeeper for the technology that governments, militaries and corporations increasingly need.

The Business Edge And Its Fragility

Anthropic’s commercial advantage comes from focus. Coding models require step-by-step reasoning, tool use and reliability, all of which are useful for broader AI agents that can work alongside employees. That makes the company’s lead in software engineering especially valuable. If AI agents move from helping coders to handling procurement, HR, finance or customer operations, Anthropic could challenge the software-as-a-service companies that currently own those workflows.

But the advantage may not last. OpenAI has moved harder into coding, Google has deep enterprise software relationships, and model performance can converge quickly as rivals copy each other’s advances. Anthropic may be ahead, but frontier AI is a market where lead times can shrink fast. The article also notes that Anthropic’s newer models use more tokens per query, which makes pricing sensitive. If OpenAI starts a price war, Anthropic’s revenue growth could come under pressure.

This is why the government’s intervention is commercially dangerous. Investors looking at Anthropic’s IPO will not only ask whether its models are better. They will ask whether the company can keep selling them without sudden political disruption. Customers will ask the same question. A model can be technically superior and still be risky to depend on if access can be suspended overnight.

The Regulatory Vacuum

The deepest problem is that both sides have a defensible concern, but there is no stable rulebook for resolving it. Anthropic wants to release useful, powerful models to customers, while preventing obvious abuses. The national-security apparatus wants control over systems that may help with hacking, cyber defence, weapons development or other strategic capabilities. Neither concern is frivolous.

The Trump administration, however, has spent much of its term attacking AI regulation and mocking the more cautious approach of the previous administration. That leaves it with blunt instruments when a model suddenly looks dangerous. In an earlier clash with Anthropic over military and surveillance uses, officials designated the company a supply-chain risk, making it harder for government agencies to work with one of America’s leading AI labs. In the latest dispute, they reportedly required special export licences for foreign access, pushing Anthropic toward a broad shutdown because it could not reliably police every user’s nationality.

The article’s point is not that Anthropic should be free to decide everything. Amodei himself argues that frontier models need standardized tests, audits and government oversight. But the state also needs tools that are precise enough to govern without wrecking the technology it wants America to lead. Otherwise, America risks government dominance over AI companies rather than national dominance in AI.

A Fight Neither Side Can Avoid

The clash between Anthropic and Washington is likely to continue because frontier AI is becoming too important for either side to step back. Anthropic’s leaders believe they are building systems that could transform science, labour and security, and perhaps become superhuman. The government sees the same capabilities and draws a different conclusion: no private company can stand above politics while building technology that may shape military and economic power.

That makes the article’s final lesson institutional rather than personal. The dispute is partly about Amodei’s temperament, Trump’s hostility and Anthropic’s unusual culture. But even a calmer administration and a more diplomatic company would face the same issue. AI models are advancing faster than law, and the gap leaves unelected executives making safety decisions until elected officials create a credible framework.

Anthropic’s predicament is therefore a preview of a broader frontier-AI problem. The firms building the strongest models want speed, customers and technical discretion. Governments want security, leverage and control. Until Congress or another durable authority writes clear rules, each new model release may become a test not only of machine intelligence, but of state power over the companies that produce it.