It's been one of those weeks where the AI industry reminds you it's still a contact sport. Three stories dropped that you actually need to know about, and they all connect in ways that matter.
OpenAI's GPT-5.6 Comes in Three Flavors
OpenAI previewed GPT-5.6 and it comes in three tiers: Sol for the hard problems, Terra for everyday use, and Luna for fast, cheap volume work. None of them are publicly available yet. Both Sol and Terra are sitting in Washington's new frontier AI review process, which is its own story.
The tiered model approach makes sense. Not every task needs the most capable model, and burning compute on a Sol-level answer when Luna would've handled it is just waste. OpenAI is clearly thinking about margins now that the VC honeymoon is over.
The government review angle is the part people are glossing over. Washington is now formally reviewing frontier model releases before they go public. We are watching the regulatory framework get built in real time, and it's going to slow things down for everyone.
Anthropic Says Alibaba Stole 29 Million Conversations
Anthropic filed a lawsuit claiming Alibaba stole 29 million conversations with Claude. Not a few thousand test queries. Twenty-nine million. That's a deliberate, large-scale data collection operation if the allegation holds up.
Training data is the real asset in AI right now. Claude conversation data specifically would be valuable to anyone building a competing model. You'd get real user queries, real responses, edge cases, and patterns that took Anthropic years and hundreds of millions of dollars to develop.
If this is true, it's one of the most brazen IP grabs in tech history. The scale makes it impossible to argue accident or gray area. I'll be watching how this plays out in court. China-US AI competition just got an official legal front.
The US Government Is Now Officially Fighting With AI Labs
The US government partially lifted a Claude Mythos 5 export control ban for critical infrastructure defenders. Which means there was a ban in place to begin with. The Anthropic-DOD dispute is being called one of the most consequential legal confrontations between an AI lab and the federal government we've seen.
A year ago nobody was talking about export controls on AI models the way they talk about chip export controls. Now government agencies are fighting with AI companies over who gets to use which model, where, and for what purpose.
This is going to get more complicated before it gets simpler. The government wants control. AI labs want to sell to everyone. National security arguments cut both ways. The labs that learn to work with Washington instead of against it will have a real structural advantage over the next few years.
This week was a reminder that AI is not just a technology race anymore. It's a legal, political, and geopolitical one too. Build accordingly.