The company unveiled Claude Sonnet 5, calling it its most agentic Sonnet yet. In plain English, it can independently plan tasks, use tools like browsers and terminals, and work through complex problems with much less hand-holding. Anthropic says it now performs close to Opus 4.8 on reasoning, coding, and agentic tasks while costing roughly 60% less through its introductory API pricing. Early partners also reported that Sonnet 5 is much better at finishing long, multi-step workflows and checking its own work without being prompted.

Then came the plot twist.

Just days after the US government pulled Fable 5 and Mythos 5 offline over national security concerns, Anthropic announced the export restrictions had been lifted. Fable 5 returns globally today, but there are still a few catches.

Through July 7, Fable 5 will only count toward up to 50% of weekly usage limits on eligible paid plans. After that, it moves to a usage-credit model, meaning heavy users will have to pay separately for additional access. Mythos 5 remains much more restricted and is only returning for approved US organizations.

The models also aren't returning exactly as before. Anthropic says new safety classifiers now block the jailbreak technique that triggered the export dispute in more than 99% of tested cases. The tradeoff is that some perfectly normal coding or debugging requests may temporarily fall back to Opus 4.8 while the classifiers are refined.

A week ago, it looked like Anthropic's most powerful models might disappear indefinitely. Today they're back online. That's a remarkably fast turnaround for what started as one of the biggest AI policy stories of the year.

Researchers discovered hidden code inside Claude Code that quietly detected certain China-related proxies and environments before subtly modifying the system prompt sent back to Anthropic's servers. The changes weren't visible to users and were designed to help identify attempts to bypass regional restrictions or resell access to the model.

To be clear, there is no evidence the software was secretly collecting user files or behaving like traditional spyware. Anthropic says it was an old internal experiment that will be removed in an upcoming release.

Even so, many developers weren't happy.

Claude Code is a tool that can access your codebase, run terminal commands, and interact with your computer. For software with that level of access, users expect complete transparency. The biggest criticism wasn't the goal of preventing abuse. It was that the feature existed at all without anyone knowing.

After spending the past week defending its models in Washington, Anthropic now faces a different challenge: rebuilding trust with the developers who use its products every day.

In less than two weeks, the company has navigated government export restrictions, brought back its flagship model, launched Sonnet 5, and hinted that even more powerful Mythos models have already finished training internally.

The AI race isn't slowing down. If anything, the last ten days proved just how quickly the leaderboard (and the headlines) can change.