Today's Overview
This week belongs to the model labs. OpenAI released GPT-5.5, positioned as "a new class of intelligence for real work." DeepSeek answered within hours with V4, open-sourcing 1.6 trillion parameters under MIT. Anthropic's Claude Code just marked one year of dominance in AI-assisted coding. The pattern is now clear: capability exploration is still the game, and the builders willing to spend the most tokens discover the next breakthroughs first.
What Actually Changed
GPT-5.5 doesn't just score higher on benchmarks. It's faster-20 minutes instead of 33 for complex tasks. It runs natively on NVIDIA GB200/300 systems. More importantly, it's being bundled into Codex as a full agent workspace with browser control, spreadsheet integration, and auto-review. Meanwhile, DeepSeek's aggressive pricing ($0.14 per million tokens for Flash, $1.74 for Pro) forces everyone's hand. When you can get 1M context window and MIT-licensed weights for a fraction of closed-model costs, the economics of agent infrastructure shift immediately.
Robotics is following a different curve. Pudu raised $150 million to shift from service robots into industrial delivery and logistics-a category where embodied AI meets real warehouse constraints. GM's Mikell Taylor is leading a "Women in Robotics Breakfast" at the Robotics Summit in May, a signal that the industry knows it has a 19% female engineering problem and isn't hiding from it. These aren't splashy model releases. They're the quiet infrastructure bets that turn research into $1.5 billion valuations.
The Real Constraint Now
Google Cloud CEO Thomas Kurian gave the clearest framing yet: agents don't just need smart models. They need identity management, memory systems that persist across steps, integration with company data through knowledge catalogs, and cybersecurity that matches the speed of AI attacks. This is why enterprises are choosing platforms over pure models. A startup can't bolt these pieces together faster than Google, Anthropic, or OpenAI shipping them as integrated products. What matters now is which platform builders actually choose to build on-and stickiness is winning over theoretical superiority. Claude Code still owns coding agents despite Codex being competitive, because first magical product experiences create lock-in that superior alternatives can't easily break.
For builders, the implication is simple: the token-maxing phase continues. Spend more, explore faster, because the companies living at the edge of capability spend the most and find the next pattern first. But the infrastructure to ship those discoveries-to move from experiment to production-is where the real margin lives. That's why robotics companies raising capital aren't just buying compute. They're building the systems that turn compute into deployed units that work in the real world.