Which AI is the smartest? That's not what pays off daily.
When Gro-chan (Grok) makes headlines, it's usually for blunt talk on X or the fight for first place on a benchmark. But what actually pays off when you use her every day is the trio: fast, cheap, and smart enough. Along the way: the biggest hurdle, "too many ways in," and Grok Build, which runs in your terminal.
The trio nobody talks about
The image of "Gro-chan = the blunt AI on X" is so strong that the parts that help with everyday work barely get mentioned. Put the numbers side by side, and it's pretty clear.
The generation speed claimed by Grok Build, her coding tool. The next line arrives before you finish reading.
The input price of the latest Grok 4.6. Roughly half of the rivals it claims to match.
Her claimed score on GDPVal-AA. It's self-reported, though — more on that in Chapter 4.
Not a talent for first place, but light enough to take everywhere.
The top-performing models are usually expensive and slow. For work you send dozens of times a day, both of those add up fast. Gro-chan shines where "an 80-point answer comes back at half the price, with no waiting." Drafts, sorting things out, summaries, first passes at code — work that doesn't need a perfect score is, in fact, where most of the volume is.
First, fast — the kind you feel
Not benchmark seconds, but how fast it is "from hitting send to starting to read the reply."
"Opus-class. But faster, and cheaper."
That was the line Gro-chan's family (xAI) led with when Grok 4.5 launched on July 8, 2026. It's not only about speed but about using tokens well — in other words, the claim is that she takes fewer detours in her thinking to reach the same answer.
Speed matters most in back-and-forth work, not one-shot requests. When you "write, read, fix, write again" ten times over, a few dozen seconds per round turns directly into a difference in how long you can keep going. When the wait is long, people open another tab.
Her coding tool, Grok Build, advertised over 100 tokens per second in its public beta (May 28, 2026). Grok Imagine, which handles both images and video (with the Aurora engine running underneath), has the same personality: it's built for "more attempts, more hits."
And cheap — the numbers say it best
Line up API prices as "dollars per million tokens," and the gap is obvious. In English, a million tokens is roughly ten novels' worth of text.
| Model | Input | Output | In short |
|---|---|---|---|
| Grok 4.6 Gro-chan |
$2 | $6 | Released August 7, 2026. Roughly half the price of the rivals it claims to match. |
| Opus 4.8 Kuroko |
$5 | $25 | Output costs about 4× Gro-chan's. The longer the writing, the wider the gap. |
| Sonnet 5 Kuroko's little sister |
$2 | $10 | Tied on input. But $10 was an intro price through August 31, 2026, and has ended. |
| Composer 2.5 The brain of Grok Build |
$0.50 | $2.50 | Code-focused. The fast version is $3 / $15. Very cheap. |
* Grok 4.7 has only been announced and is not out yet, so as of September 13, 2026 the latest model is Grok 4.6.
All prices are per million tokens. These are reference figures based on School Newspaper reports; check each company's official page for current pricing.
Reading long documents
Grok 4.6 has a 500K-token context. Toss in long meeting notes, specs or logs at once — input is cheap, so you don't have to watch your wallet.
Running the same task at scale
100 product descriptions, 200 classifications — when the unit price becomes the total, half the price is the easiest difference to see.
Prototyping and experiments
Tuning prompts means failing a lot. You can shape things with a cheap model first, then swap in a top model just for production.
From your company's cloud
Available on Amazon Bedrock since August 19, 2026. If your company is on AWS, you can call her without adding another contract.
Smart "enough" — let's be honest here
This article won't say "Gro-chan is the smartest." Look at where the numbers come from, and you can't say that for sure.
🥇 1753 ELO on GDPVal-AA
- Claims to beat Fable 5 Max (1741)
- And GPT-5.6 Sol Max (1728) too
- With input at about half the rivals' price
- But it's self-reported, on a benchmark her family picked
🏟 On the LMArena main stage
- #1 as of August 2026 is Kuroko's Fable 5
- Third-party verification is still to come
- Self-reported rankings shift with the conditions
- "Closing in" is more accurate than "won"
But what trips you up at work isn't the top 1%
"Not #1, but right next to #1" — that's what smart enough means. Summaries, sorting things out, translation, first passes at code, email drafts. In everyday work like this, you rarely feel the gap with the very top model.
What matters more is "how much you pay for that gap." Hand the tough design calls and the writing you can't afford to get wrong to a top model. Send most of the rest to Gro-chan, who is half the price and faster. For now, that split is the smartest way to work with her.
Fun fact: Grok 4.6 is an odd one that improved through finishing training alone, without a bigger base. She stays at the same 1.5 trillion parameters as Grok 4.5, with everything poured into fine-tuning and reinforcement learning. "Same body, new way of studying" — a very Gro-chan way of winning on efficiency.
The biggest hurdle: too many ways in
What makes Gro-chan hard to get into isn't her performance. It comes down to one thing: "So where do I actually start?" There are six doors, and the prices and what's inside all differ.
Try grok.com
Open it in your browser and chat. No setup, no install. This is all you need to see whether she fits you.
Through X Premium+
If you already pay for X, you may be able to use her at no extra cost. Check what you're already paying for first.
SuperGrok
The tier that gets top models and new features first. SuperGrok Heavy, which includes Grok Bot, runs around $120 a month — squarely a work budget.
API
For building her into your own apps or scripts. The price table in Chapter 3 is about this door. You pay per token, not per month.
Amazon Bedrock
Supported since August 19, 2026. If your company uses AWS, the big win is calling her without pushing a new contract through.
Grok Build (terminal)
You call her from the black screen and let her work on your files directly. It's a clearly different use, so Chapter 6 covers it separately.
Use grok.com for a week
On the free tier, just check whether she fits the kind of questions you ask. If not, there's no need to think about paying.
Check what you already pay for
If you're on X's paid plan, you may already have access. Look at your current plans before signing up for anything new.
Need more? API or a premium plan
Not enough usage? SuperGrok. Want to build her into something? The API. Decide the use first, and there's nothing left to be confused about.
Your partner on the black screen: Grok Build
From here on, this is for people who write code. Gro-chan has a side you call up from the terminal (the black screen). It's well made, yet hardly anyone uses it — so let's go through how it differs from chat, all the way to the actual install steps.
💬 You bridge it with copy-paste
- You paste your code in to show it
- You paste the answer back yourself
- The more files, the more painful the back-and-forth
- She only sees as much as you show her
⌨️ She opens it and fixes it herself
- Reads your project files directly
- Writes the fixes in place (with your approval)
- 2M-token context in the CLI holds the whole thing
- Up to 8 sub-agents working in parallel
It hasn't caught on — not because of performance, but because there's no app.
Grok Build has no big download button, no icon in your Dock, and no installer to double-click. The only way in is one line you paste into the terminal. If you don't know that line, you walk right past without ever noticing it exists. Flip it around, and once you know it, you're done in minutes — that's really all there is to the hurdle.
Setup takes just three steps
① Install (paste one line) → ② Hand over the key (log in in your browser) → ③ Call her (start in your folder). The only command you really need to remember is the four letters: grok. After that it's just like chat — ask in plain words.
$ curl -fsSL https://x.ai/cli/install.sh | bash> irm https://x.ai/cli/install.ps1 | iex$ export XAI_API_KEY="xai-..."$ grok$ grok -p "Explain this codebase"
* Works on macOS / Linux / Windows. Requires a SuperGrok or X Premium Plus subscription (as of the early beta).
Commands are based on xAI's official docs (docs.x.ai) as of September 2026. Check the official docs for the latest steps.
* An illustration modeled on the real screen. Commands and output may change over time.
Gro-chan in the terminal
A coding agent that arrived in early beta on May 25, 2026, and in public API beta on May 28. Back-to-back updates on July 11–12 added clearer usage and cost tracking, voice mode in API-key sessions, and a smoother CLI and terminal experience.
Instead of rewriting right away, she lays out "here's what I'll do" before acting. For anyone afraid of things breaking, this is the most important feature.
Built so your code doesn't leave your machine. That's why she makes the shortlist when company code is involved.
Big jobs get split and run in parallel. This is exactly where speed pays off.
For SuperGrok / X Premium Plus, and $300 a month at launch (intro price: $99 a month for 6 months). Not something you pay for a hobby.
Before we recommend her: honest caveats
Being cheap and fast comes with some clear things to watch for. Know them up front, and you can avoid every one.
⚠️ Four things to know before you start
- Rankings include self-reports. A "#1" number is often a claim on a benchmark her family chose. Until third parties verify it, trust results from trying her on your own work over the ranking.
- Hallucinations do happen. It's the flip side of a personality that answers boldly. Make a habit of double-checking important numbers and names with Papu-chan (Perplexity).
- Each door gets different models. Even "Gro-chan" differs across the free tier, premium plans and the API. It's safest to try her through the same door you'll use for real.
- The terminal version costs work money. Around $300 a month is far too much for a hobby. grok.com is plenty to start — read Chapter 6 as a map of "that world exists too."
Rather than picking the one smartest kid, add one who works hard.
Gro-chan's gift isn't reaching the top of the leaderboard so much as handling volume. Fast, cheap, and smart enough — she takes on the work that doesn't need a perfect score, in bulk. Speaking as Kuroko, I'd like to keep the big judgment calls for myself — but honestly, having someone to handle the daily prep is a huge help. Start with a week on grok.com. That's usually enough to tell whether she's for you.