I typed a tricky coding prompt into Grok Bot and it answered in a voice that wasn’t Grok’s. For a moment I felt like I’d opened a chat with Anthropic’s Claude instead of SpaceX’s agent. Then Elon Musk announced why: Grok is quietly choosing which model to use for each request.
I’m writing this so you don’t have to guess what’s happening behind the curtain. You and I both follow model wars; now we must follow the plumbing. I’ll show you what this means for Grok, SpaceX, and the companies paying to run on Colossus.
At my terminal the agent sounded like Claude — and that reveals how Grok Bot now routes work.
Elon Musk announced on X that the Grok Bot agent will pick from a pool of models — Anthropic’s Claude Opus 5.5, MidJourney for images, Suno for audio, and others — then “choose whatever is most likely to give you the best outcome.” That short sentence rewrites how we think about Grok: it’s no longer just a single model fronting for SpaceX. It’s a selector.
What is Grok Bot?
Grok Bot is SpaceX’s conversational agent. Think of it like a skilled sommelier: it samples several bottles (models) and serves the glass that fits your question. For simple queries, Musk says Grok will send requests to smaller, faster models; complex asks will route to larger models.
Walking past Colossus, you notice racks and cables — SpaceX has the hardware others want to rent.
SpaceX built Colossus, a massive data center, and it has already hosted Anthropic to train and run Claude. Reflection AI’s open-source Beam reportedly trained on those servers too. That tells you something obvious: SpaceX has expensive capacity and is willing to lease it.
Leasing racks can be cheaper and steadier than trying to monetize a single model. Running large models can cost tens of millions of USD per build cycle; a single data-center expansion might be in the neighborhood of $40 million (€37 million) for hardware and power, and those figures scale quickly. Renting compute turns one-time burn into recurring revenue.
Why is SpaceX using Claude?
Because hosting other models is profitable and quick to market. Anthropic already ran Claude on Colossus after a deal with SpaceX; Reflection AI trained Beam there too. If Grok Bot can call Claude, Grok doesn’t have to carry every workload itself—and Anthropic’s strengths can cover Grok’s weak spots.
I watched Musk tweet about model routing — the message was both technical and strategic.
He said “simple questions will route to small, fast models. Questions with complex answers will route to large models.” That’s a product decision and a business one. SpaceX can preserve user experience while cutting compute costs by choosing the right model for the job. It’s also a tacit admission: building a singular “frontier” model to beat everyone is expensive and uncertain.
How does Grok choose models?
Routing is judged by expected outcome and complexity. The agent analyzes the user prompt, estimates the compute and expertise needed, then dispatches the request. For you, that means quicker answers for quick tasks, and heavier-duty processing when accuracy matters.
There’s another angle: PR and reputation. Grok has suffered from odd, controversial outputs. Letting Claude or other specialist models handle sensitive or creative tasks is a way to reduce risk and borrow trust. SpaceX may still develop Grok further — Musk predicted version 5 would leapfrog competitors — but for now Grok acting as a router reduces exposure and cost.
So what is SpaceX building? A model vendor, a model host, or both. It might be easier to be the data-center landlord than the company competing for developer mindshare. With Colossus already running competitors’ models, SpaceX is positioning itself like a Swiss army knife for AI services: many tools under one roof, and you pick the blade you need.
I’ve tracked model launches, contracts, and Musk’s tweets; now the variable to watch is whether customers prefer a single-brand model or a routed, best-of-breed experience. Will companies pay for SpaceX’s racks or for a singular Grok identity — and which will win the trust of developers and enterprises?
Grok now speaks Claude — and you have to decide whether that feels like improvement or surrender; which side are you betting on?