THIS WEEK IN AI: Grok Bot (and 4.6), AI Watermarks, DeepSeek Strategy, Selling the Lakers

The week's biggest surprise came from an unexpected direction: SpaceX AI released Grok 4.6, and it is now arguably the best frontier model available — depending on what you care about. On the AA Intelligence Index, a benchmark measuring general intelligence across models, Grok 4.6 High scores at the level of OpenAI's GPT-5.6 Sol Max and within a single point of Anthropic's Fable 5 Max. The kicker is cost: Grok 4.6 is 5x cheaper to run than Sol and 8.5x cheaper than Fable. For companies spending tens or hundreds of millions on frontier models, that matters.

How SpaceX AI Got Back to the Frontier

SpaceX AI had lost the frontier race for a while. Elon's engineering team made some wrong moves and fell behind Anthropic, OpenAI, and Google. But Grok 4.6 — released just three weeks after Grok 4.5 — represents a return. The model achieves frontier intelligence through a multi-agent architecture: it spins up multiple agents to solve a query, a technique Meta and Chinese open-source models also use. But SpaceX AI has two decisive advantages. First, they have enormous amounts of proprietary data — including all of SpaceX's rocket science data, which no model has ever been trained on. The Starship rocket was built 100% by humans with no AI, so all those schematics, science, and engineering map