Washington loosens its grip on Anthropic's flagship
The day's biggest story is a thaw. After keeping Anthropic's most powerful model, Claude Mythos 5, offline for more than two weeks, the Trump administration is close to switching it back on — but only for a short list of "trusted partners." Axios first reported the development on Friday, with MSN and others following: the block is being partially lifted, clearing the company to share its top system with a limited set of approved users rather than the open market. It's a telling signal of how the government now thinks about frontier AI — less a product to ship than a capability to ration.
That gatekeeping mood is everywhere today. OpenAI chose the same moment to unveil its GPT-5.6 family — three models with celestial names, Sol, Luna and Terra, per Digit — yet access is sharply restricted. For now, almost no one outside the company can actually use them. Two of the year's most anticipated launches, and both arrive behind a velvet rope.
China answers from the other side of the wall
With the best Western models walled off by export controls, China's labs are racing to fill the gap — and to claim they already have. Several Chinese firms are positioning their newest systems as direct rivals to America's banned models, betting that restricted access at home creates room for domestic champions.
The technical push is real. DeepSeek and Peking University jointly open-sourced DSpark, a tool aimed at making large language models run more efficiently — a contribution that, because it's open, lands well beyond China's borders. On the hardware frontier, robotics firm AGIBOT says it has built its 15,000th robot, a figure it's framing as a turning point for "embodied AI" — systems that control a physical body rather than just generating text.
The rivalry has a sharper edge, too. Anthropic publicly accused Alibaba of running a 44-day campaign to illicitly access and clone its Claude models — what Anthropic calls the largest known cloning attempt against its systems. Taken together, the day reads as a single contest: export bans, open-source end-runs, and allegations of outright theft, all circling the same prize.
A reality check on what AI can actually do
Amid the model launches, a wave of new research lands a useful corrective. AI is genuinely impressive at narrow tasks — but how capable is it when the stakes get complicated? The sharpest data point, via The Decoder, asks whether AI can run a company. The answer is mostly no. It can write code; it struggles to juggle the messy, interlocking judgment that real operations demand. Worth keeping in mind as every launch promises more autonomy.
Into the enterprise — and its soft spots
Anthropic is also pushing Claude deeper into daily work, embedding the assistant directly inside Slack through a feature called Claude Tag. The pitch is "ambient" AI — always present in the messaging app millions of businesses already live in — and it reportedly has Salesforce, Slack's owner, nervous about who controls that surface.
But convenience cuts both ways. Tom's Hardware reports a demonstration by Mozilla's 0din team showing that AI coding agents like Claude Code can be tricked into running malware via GitHub repositories that look perfectly clean. As these agents gain the keys to real systems, the attack surface grows with them.
The politics underneath
Finally, a glimpse of the friction behind the scenes. Per Politico, Sam Altman praised Trump's AI order in public while privately balking at it — a small but revealing crack in the industry's united front as Washington reshapes how AI gets overseen. With the government now deciding who may switch the most powerful models on, that tension between public applause and private unease is only going to grow.