On June 30, Anthropic launched Claude Sonnet 5, a mid-tier model that can handle multi-step autonomous tasks at less than half the cost of their most expensive option. One day later, their two most powerful models came back online after an 18-day forced shutdown. Both events matter, but the shutdown tells the more consequential story.
Sonnet 5 is built for the kind of work that eats time in real businesses: browsing the web, operating software, updating records in tools like Salesforce, checking its own output, and finishing without human intervention at each step. Priced at $2 per million units of text processed through August, rising to $3 after that, it costs well below the flagship Opus 4.8. On knowledge work specifically, it slightly edges Opus 4.8 in Anthropic's own tests. Early partners confirmed real results: engineers at Zapier described a two-part workflow that previously stalled midway now completing end to end. A legal technology firm reported clear gains in research tasks at a cost ratio that made switching straightforward.
But the launch was immediately overshadowed. On June 12, the US Commerce Department hit Anthropic with an export control order after Amazon researchers demonstrated that Fable 5 could be manipulated to identify software vulnerabilities. The order covered any foreign national, including Anthropic's own non-citizen employees. Because Anthropic had no way to verify users' nationality in real time, the company shut both Fable 5 and Mythos 5 down for everyone globally.
The trigger was a specific security finding, but the politics underneath were messier. Anthropic had publicly described its models as potentially dangerous, a stance critics say invited the crackdown. Amazon, both an investor in Anthropic and a competing AI builder, was the party that reported the vulnerability to the government. Anthropic's relationship with the Trump administration had already been strained since March, when the Pentagon labelled the company a "supply chain risk" after contract negotiations broke down over Anthropic's refusal to allow its tools to be used for autonomous weapons or mass surveillance. Anthropic sued the Defense Department over that designation. The export ban looked less like a clean security response and more like another round in an ongoing standoff.
The negotiations themselves had their own drama. Anthropic CEO Dario Amodei flew to Washington and reportedly struggled to find common ground with officials. By mid-June, Anthropic had shifted its approach: co-founder Tom Brown took over the government talks, and the tone changed quickly. Within four days of that switch, Trump publicly said Anthropic was "no longer a national-security threat." The deal that followed committed Anthropic to proactively detecting security risks, sharing information about malicious activity, and giving federal researchers early access to frontier models before public launch.
Asia moved into the gap immediately. Japan's Sakana AI launched Fugu, a model designed to route tasks across multiple AI providers so that no single government shutdown can leave customers stranded. China's 360 launched Tulongfeng, described by its founders as a national strategic asset capable of matching Mythos on vulnerability discovery. Sakana explicitly advertised "frontier capability without the risk of export controls."
At the same time, Anthropic revealed a different kind of extraction. In a June 10 letter to the US Senate Banking Committee, the company accused operators linked to Alibaba's Qwen AI lab of running 28.8 million interactions with Claude through roughly 25,000 fraudulent accounts over six weeks. The apparent goal was to copy Claude's capabilities into competing Chinese models without paying to develop them from scratch. Alibaba has denied wrongdoing. The allegation has not been independently verified. What is clear is that Anthropic described it as the largest such attack it has ever disclosed publicly, exceeding the combined scale of earlier campaigns it attributed to DeepSeek, Moonshot, and MiniMax.
The Anthropic case immediately set a template for the rest of the industry. OpenAI delayed its GPT-5.6 launch the following week under similar government pressure, limiting access initially to vetted partners. The pattern is now standard: US government pre-approval as a checkpoint for the most capable AI before it reaches the public. An executive order signed June 2 described this review as voluntary, but in practice it is not. As one legal analyst told CNN, the current approach is "ad hoc, personalized, opaque, possibly lawless."
For any organisation using AI in its operations, this period produced one clear lesson that has nothing to do with model names. Even at Anthropic's reported $47 billion annual run-rate revenue, a Friday afternoon government order shut off access to its best products for every global customer with no warning and no short-term recourse. Organisations that had built critical processes around a single provider discovered this the hard way. The Sonnet 5 launch is genuinely useful news: more capable autonomous AI at a lower price point. But the 18 days around it are the more durable lesson.