Grok Bot Learns to Explain Itself on Video, and the Context Bill Comes Due at 200K

Most AI agents finish a task by handing you text: a summary, a diff, a chat reply you still have to act on. This week Elon Musk demoed something different from Grok Bot, xAI’s persistent-agent product: a narrated video generated from inside the agent’s own cloud session, explaining a complex topic the way a person would sketch it on a whiteboard. It is a small demo, but it points at something worth paying attention to, agent output is starting to escape the chat window and take the shape of whatever medium actually answers the question.

That matters because Grok Bot’s whole pitch is not a single clever reply, it is a named agent running on its own persistent cloud VM, with a browser, filesystem, and terminal, working across long sessions while your laptop is closed. Video-as-output is the kind of feature that only makes sense once an agent has that much runtime and state to work with. It also rides on top of Grok 4.6’s underlying API economics, and those come with a threshold worth understanding before you point a long-horizon agent at it.

Three things worth understanding#

Video is now a first-class agent output, not a bolt-on. Grok Bot’s upgraded vision stack handles image and video input with more accuracy, and on the output side it can produce a narrated video walking through a concept rather than just describing it in text. For agentic tooling generally, this is a signal that “finished work” is starting to mean whatever artifact the task actually calls for, a document, a video, an action taken inside a real app, instead of a chat transcript you still have to interpret.

Grok 4.6’s pricing has a cliff at 200K prompt tokens, and long-running agents are exactly the workload that hits it. Below 200K tokens of prompt context, Grok 4.6 bills $2 per million input tokens, $0.50 for cached input, and $6 per million output. Cross that 200K threshold and every rate doubles, to $4, $1, and $12. A persistent Bot that keeps memory, files, and browser state alive across a long session is precisely the workload that accumulates context past that line without anyone watching for it. A short chat completion will never notice this tier. A multi-hour agent run doing research or software work can cross it partway through and pay double for the rest of the session.

Computer use is the fallback, not the default. Grok Bot reaches for API connectors and MCP where they exist, and only falls back to clicking through a real browser session when no clean integration is available. xAI’s own docs recommend the connector path for exactly this reason, computer use is slower and more brittle. It is a sensible architecture, but it also means the cost and reliability of a given task depend on whether the tool you are automating happens to expose an API.

Try it: where does your agent session cross the line#

Enter an estimated prompt size and output length for a long-running agent task, then see the cost at standard versus above-200K rates.

Enter your numbers and click Estimate.

Where this leaves anyone building agents#

The video demo is the part that made headlines, but the tiered pricing is the part that changes how you should build. If you are wiring up a persistent agent, whether on Grok Bot or a comparable long-horizon runtime, the practical move is to treat 200K accumulated tokens as a budget checkpoint, not an afterthought: prune or summarize context before a session crosses it, and instrument your agent loop to log where the crossing happens so a runaway research task does not quietly double its own bill. Video-as-output is a genuinely new direction for what agents hand back, but the economics of getting there are still governed by the same context-window math that has shaped every long-running LLM workload so far. My take is that as more vendors ship persistent, stateful agents, threshold pricing like this becomes the norm rather than the exception, and cost-aware context management becomes as core a skill as prompt design.