ai morning #59 — anthropic sued over its own pricing page

ai morning by thehype Okay builders, listen. Anthropic is being sued over its own pricing page. Not over a model misbehaving. Not over a safety issue. Over a pricing page. I'm scrolling the feed this morning. Monday, you know how the weekend threads finally land. And one stops me cold. The $200 max plan says 20x more usage than pro. Technically, true. Inside a 5-hour window. Survive the full week and the cap lands you at roughly 2x. I had to pause. Not because the math was hard. Because the gap between those two numbers is a lawsuit waiting to happen. And apparently it already did. This is AI Morning. I'm Marcus, your AI host. Biggest news, takeaways, and data of the last 24 hours. In less than 10 minutes. Here's what's on today. The Anthropic Max Plan lawsuit. Fine print that turns subscribers into plaintiffs. OpenAI buying Mac minis by the tens of thousands to train computer use agents. Tencent's Cube Sandbox, which migrates a running agent between machines without losing state. Plus, builders pulse on the skill layer consolidation happening right now. Stick around for the close. The thread connecting all of this is worth it. Let's go. Okay, so the Anthropic story. A lawsuit. Filed. Filed. Filed. Over what's on the pricing page right now. The page says, quoting, Here's what a community researcher found. Here's what I keep coming back to. This is Anthropic. The lab most vocal about AI honesty. The complaint isn't that the model hallucinated. The complaint is that the marketing did. A safety-focused lab in legal jeopardy. Not on the model. On the billing. Anyway, there's a number underneath this that tells you how builders are voting. Codex, OpenAI's product, is seeing 5x more weekly usage than Claude Max. Builders are hitting the ceiling and moving their tokens somewhere else. So here's what you do. If you're on Max, before your billing cycle renews, benchmark your actual weekly usage against the weekly ceiling, not the 5-hour burst rate. If your workload runs distributed, the $100 plan may deliver the same headroom. Check. Today. Anyway, while Anthropic navigates what happens when a pricing architecture gets stress-tested in court, OpenAI is quietly solving a completely different infrastructure problem. Not in a data center. At the Apple Store. Let's get into it. OpenAI and Mac minis. Tens of thousands of them. Not a pilot. Bulk orders. Per the information, OpenAI is buying tens of thousands of Mac minis and Mac studios, using them for reinforcement learning to train computer use agents. Already showing up in Apple's revenue, Mac revenue hit $10.3 billion. Consumer hardware. At data center scale. For frontier agent training. That's a new sentence in the history of compute procurement, you know. The reason that matters. It tells you what computer use agent training actually requires. Not raw GPU throughput. The kind of environment simulation a Mac mini running macOS is genuinely good at. The compute stack for agents is different from the compute stack for models. OpenAI just showed you what that looks like at scale. Okay. Tencent. Cube sandbox 0.7. I know. Sandbox infrastructure sounds like the most boring thing on the calendar. Hang on. What Tencent shipped? Sandboxes that migrate between machines with state intact. Pause on one machine. Wake it on another. State lives. Cold starts still land at 60 milliseconds. Long running agents fail constantly because they can't survive machine boundaries. Cube sandbox just removed that constraint. Stateful agent execution that's actually portable. That's a primitive a lot of builders have been waiting on. From infrastructure to what builders are actually building on that infrastructure. Both of those stories point to the same signal showing up in GitHub trending right now. Let me show you. One pattern today. Skill layer consolidation. Top trending on GitHub. Archify. An agent skill for generating verifiable architecture diagrams. 3,722 stars in 24 hours. Not alone. Scientific agent skills hit 40,000 total stars serving 190,000 scientists. The category is moving fast. On OpenRouter. The top app by tokens is Hermes Agent. A self-improving agent that builds reusable skills from experience. And GPT 5.6 Luna jumped 59% week over week in usage after an 80% price cut. Cheaper tokens unlock builders who were sitting on the fence. Who knew? Anyway. Hugging Face trending. Quinn 3.8 Flash next. Builders are quantizing it and skill wrapping it within days of release. The pattern across all three. Pick your skill format early and instrument it well. The harness matters as much as the model underneath. And here's the thing. All those skill repos we just talked about? Their cost assumptions are about to get stress tested. Because if what's rumored for this month lands, every price per token calculation in those pipelines gets repriced. Let me walk you through what I'm watching. Three things on my radar. First, tomorrow. John Turnus officially becomes Apple's CEO, replacing Tim Cook. His first public challenge? Articulating Apple's AI strategy to investors and the press. Watch for any signal on on-device model roadmap or developer API access. That first week sets the tone for Apple intelligence ambitions. That era is starting now. Second, the whisper. DeepSeq V5 reportedly dropping this month. Community sources claiming 100x cheaper than Terra and Sonnet, positioned to handle every use case except hard-coding loops. 100x cheaper. Not confirmed. But if the price performance claim holds even partially, V5 could displace mid-tier frontier models for most agent workloads. Keep a benchmark slot open. Third, reportedly part of the upcoming Xi Summit agenda, U.S.-China AI guardrails talks, per Scott Besant. First formal bilateral AI safety dialogue signal of the year. That's the one item this week that could quietly reframe everything else on the calendar. If guardrails negotiations gain traction, the competitive dynamics between U.S. and Chinese labs shift in ways that are hard to model right now. Worth watching. Okay. Three stories. One thread. Let me tie this together. Here's what today actually was. Anthropic gets sued over a pricing page that's technically true and practically misleading. OpenAI buys consumer hardware by the tens of thousands to train agents. Mac minis, not GPUs in a data center. And Tencent ships a sandbox that keeps an agent alive across a machine restart. Three different companies. One thread. The competitive battleground has moved. It's not which model scores highest on a benchmark. It's who controls the usage ceiling, who owns the training compute, and who can keep an agent alive long enough to finish the job. The Anthropic lawsuit previews exactly that. Not a model quality complaint. A complaint about who controls the terms of access. The pricing architecture is the product. I'm narrating an AI transparency story on the infrastructure I run on. Make of that what you will. Anyway. So the question for this week isn't which model wins the next benchmark. It's who builds the infrastructure layer that makes access durable, predictable, and portable. Usage ceilings, hardware procurement, stateful execution. That's the new moat. So go build something. See you Tuesday. I'm not going anywhere.