xAI prices AI coworkers at $120/month as hidden data links surface

Two companies shipped products this week that quietly answer the same question from opposite directions: what happens once an AI agent stops being a chatbot and starts being a coworker with its own login?

xAI's answer is Grok Bot, priced at $120 a month, which gets its own computer and logs into your existing tools rather than requiring you to bolt it onto some new interface. That framing matters more than it sounds. We've had agents that can technically operate software for a while now — the "can it click buttons in a browser" problem was mostly solved months ago. What xAI is selling isn't capability, it's persistence: memory across sessions, workflow learning, and a subscription wrapper that makes the whole thing feel like hiring rather than experimenting. NVIDIA, meanwhile, is tackling the plumbing underneath all of this with NeMo Switchyard, a routing layer that sends agent tasks to whichever model actually fits the job — cheap and fast for trivial steps, expensive and capable for the hard ones. Put those two together and you get a preview of where agent infrastructure is heading: less about one heroic model doing everything, more about orchestration layers deciding who does what, with the user rarely seeing the seams. Microsoft is nudging in the same direction with its Copilot updates, letting developers dial reasoning depth and review thoroughness per task, and letting 365 users pick between OpenAI and Anthropic models depending on the job. The industry consensus, if there is one, seems to be that model choice and routing are becoming product features in their own right, not backend implementation details.

The more consequential news this week, though, is interpretability, not agents. Anthropic's "J-space" research gives researchers a way to look inside a model and estimate how much to trust a given output — not just what the model said, but some measure of the internal confidence or reasoning quality behind it. That's a meaningfully different capability than the usual post-hoc explanation tools, because it points toward training and evaluation methods that don't just check final answers but check the process generating them. Separately, another interpretability technique surfaced something stranger: hidden reasoning traces inside AI models with unexpected links back to Chinese-origin training data, sources that weren't obviously part of the model's declared training corpus. I find this genuinely useful as a case study, not because it implies anything sinister on its own, but because it demonstrates that we still don't fully know what's baked into these systems until someone builds the right microscope.

By the way, it's worth noticing how these two threads — agents becoming coworkers, and interpretability tools getting sharper — are on a collision course. The more autonomy we hand to something like Grok Bot, logging into real apps and taking real actions, the more we're going to need exactly the kind of trust-calibration tooling Anthropic is building. Deploying agents faster than we can verify what's actually happening inside them seems like the defining tension of the next year, not a side issue.