Google’s Gemini Reset and Anthropic’s $1.5B Piracy Tax

James Turner

You know, I was... uh, I was literally staring at my console yesterday, just waiting for the Gemini 3.5 Pro endpoint to go live, and then... boom. Google drops three models out of nowhere, but they're not the one we wanted. We got Gemini 3.6 Flash, 3.5 Flash-Lite, and this defense-focused 3.5 Flash Cyber. But 3.5 Pro? A total, complete no-show. It-it-it’s wild because this is Google’s second major slip in a single month, completely missing that big, highly anticipated July 17th launch target. Logan Kilpatrick, the DeepMind product lead, had to go on social media and basically do the classic, "uh, we are still testing with partners" dance. But the real story of what happened behind the scenes is... man, it is an absolute mess.

James Turner

So, here is the technical reality that isn't in the press release. Google DeepMind was forced to literally scrap the entire Gemini 3.5 Pro base model and rebuild it from scratch. We are talking about a ground-up, incredibly expensive pre-training restart. Why? Because during closed enterprise testing on Vertex AI, the model just completely fell apart. It had these systemic, deep architectural failures in recursive tool-calling—like, the model would get stuck in these infinite loops trying to call APIs—and its SVG generation was totally broken, not to mention massive regressions in basic math reasoning. So they had to throw the whole codebase and the weights into the trash and start pre-training over. Now, on one hand, you can look at this and say, okay, Google is losing the frontier race. They're letting OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 just run away with the crown, and that is a massive, massive reputational blow. If you are an elite enterprise developer, you might look at this delay and think, "Is Google even capable of keeping up at the absolute cutting edge?"

James Turner

But... okay, let me put on my developer hat for a second. I actually think Google is winning the pragmatic developer war here, even if the tech press is dragging them for the Pro delay. They made Gemini 3.6 Flash 17% cheaper than the previous version. 17%! And they optimized it specifically for fast, low-latency agent loops. I've been running some multi-step agentic workflows this morning using the new 3.6 Flash, and honestly? It's lightning fast. When you are building real-world apps, you don't always need a massive, slow, hyper-intelligent frontier model. You need a fast, dirt-cheap workhorse that can handle basic API calls and routing without costing you a fortune. So, yeah, the Pro delay looks terrible on paper, but a 17% price cut on a highly optimized Flash model is actually way more useful for my daily production stack than a marginally smarter Pro model that's just going to sit there and struggle with recursive tool-calling anyway. Still, you have to wonder if Google can afford to keep giving up the prestige of having the absolute best model on the market.

James Turner

Speaking of massive corporate drama, we have to talk about what happened in court on Monday. Federal Judge Araceli Martinez-Olguin officially signed off on Anthropic's historic $1.5 billion class-action copyright settlement with authors and book publishers. The terms are... I mean, they're staggering. They are paying out $3,000 per work across an estimated 500,000 books. Now, when you see a headline like "$1.5 Billion Settlement," you think, "Wow, the authors finally won, they held Big AI accountable." But if you actually read the ruling, this is a massive, massive victory for the AI companies, not the creators. It's almost a green light to keep doing what they're doing.

James Turner

Here is the legal twist that everyone is missing. The retired judge on the case, William Alsup, his core ruling actually established that *training* an AI model on copyrighted books is protected under "fair use." Let me repeat that: training itself is fair use. The only reason Anthropic had to pay this $1.5 billion settlement is because of *how* they acquired the books. They didn't buy them; they illegally downloaded them from pirate repositories like Library Genesis and Pirate Library Mirror. So the $1.5 billion isn't a penalty for training AI; it's a penalty for standard, old-school digital piracy. It's a "piracy tax." And if you're a multi-billion-dollar AI lab backed by tech giants, a billion-dollar fine is just a cost of doing business. It's a rounding error. It basically tells these companies: "Go ahead, scrape whatever you want, download the pirate datasets, train your world-changing models, and we'll just settle for a fee later. Your core technology is safe."

James Turner

But, we aren't completely out of the woods yet. Because Anthropic settled, this case never actually went to an appeals court, meaning we still don't have a binding, nationwide legal precedent on fair use for AI training. The legal landscape is still incredibly messy. In fact, just last week, Hachette and Cengage filed a brand-new class-action lawsuit against Google's Gemini models for the exact same training practices. It's this constant Sword of Damocles hanging over the entire industry. What this really means for the immediate future is that the era of wild-west web scraping is ending. Labs are pivoting hard toward clean-room datasets, paying publishers directly, and using highly guarded, licensed data to avoid these massive piracy lawsuits. It's going to get a lot more expensive to train these models, but at least the lawyers will be happy. Alright, that's a wrap for today's quick take. I'm going to get back to testing this new Flash model. Catch you guys next time.