“Grok bot” is either the @grok account that answers on X, a bot somebody wired to xAI’s Grok API, or just the Grok chat itself. None of those keeps what it makes. Inside Amelia’s Agent, Grok 4.6 is one of the models you can hand a build to: it costs $2 per million input tokens and $6 per million output tokens, it is strong at general coding with quick turns, and whatever it builds lands in a git repo you own, hosted, with the bill printed on the answer.
Search for “grok bot” and you get three unrelated answers, because three unrelated things carry the name. Sorting them out is most of the work of deciding whether Grok is the right tool for what you are trying to do.
1. The @grok bot on X
Tag @grok under a post on X and it replies in the thread. This is the thing most people mean. It is good at what a reply can be good at: summarising the thread, checking a claim, explaining a chart, being funny on purpose. It is a conversation, and a conversation is where it ends. It cannot build you anything, and it does not remember you tomorrow.
2. A bot built on the Grok API
xAI sells Grok as an API, so developers wire it into Telegram bots, Discord bots, support widgets and internal tools. The model is the same family; the bot is whatever the developer wrote around it. When somebody says “we built a Grok bot,” this is what they built, and the interesting part is the plumbing they had to add: memory, tools, a place for the output to live, and a way to stop it spending money when it gets stuck.
That plumbing is the part nobody budgets for. A model call is one line. A bot that can be trusted with a task is several thousand.
3. Grok, the model, used as a coding agent
This is the version that matters if you want something built rather than answered. Grok 4.6 is a capable general coder with quick turns, which is exactly the profile that wins on ordinary web work: a landing page, a booking form, a small app with a database behind it. Inside Amelia’s Agent it sits in the model list next to about three hundred others, priced at $2.00 per million input tokens and $6.00 per million output tokens, available from the free trial, with vision turned on so it can look at a screenshot of what it built.
You can choose it by name. Or you can leave the model on Balanced, which picks per job and climbs one rung at a time when a cheaper model cannot finish. Grok 4.6 is a rung on that ladder, between GLM 5.3 and the frontier tier. It gets the job when the job earns it, not because it is fashionable.
What we can and cannot tell you about it
We publish the measurements behind our recommendations, so this deserves a straight answer: as of early September, Grok 4.6 had six recorded production runs on our platform against 821 for the default model. That is not a verdict on Grok. It is a symptom of a loop we wrote about in the code itself: whatever the default is gathers all the evidence, so it stays the default. We added a capability floor so harder jobs are routed to models that are measurably good at coding regardless of how few samples they have, and Grok 4.6 clears it.
So the honest recommendation is this. If you already like Grok, pick it and watch the price on each answer; for page builds and small apps it will usually finish in one turn, and one turn at $2 in and $6 out is cheap. If you do not have a preference, let Balanced decide, and look at which model it chose and why. The one number that predicts your bill is turns to done, not the rate card, and that holds for Grok the same as for everything else.
What a “Grok bot” cannot do, and what fixes that
The reply bot on X and the chat cannot hand you a repository. A bot on the API can, but only after someone builds the surrounding machine. The gap between “Grok answered” and “Grok shipped” is hosting, a git history, a domain, a place to store data, a budget it cannot exceed, and a person who is asked before anything irreversible happens.
That gap is the product. Describe the thing, choose Grok 4.6 or let the platform choose, and what comes back is a running site or app with its code in your own repo and the exact charge underneath. Start from a blank prompt, or from one of the 53 templates so the first turn is “change this” rather than “make this.”
Compare Grok 4.6 with every other model Start from a template
Get the next one.
Working notes on building with AI, with the measurements and the bills attached. Sent when there is something worth sending, which is not weekly.
No spam, and one click to leave. We do not sell or share the list.
Try Grok 4.6 on a real job.
Pick it from the model list, or let Balanced pick it when the job calls for it. The price is on every answer.
Free to start · no card · every answer shows its price