"Grok Build feels like the previous generation of coding models."
That's from a Reddit post on r/grok. From someone who actually paid $99 to try SuperGrok Heavy.
Another commenter chimed in with their own story. Three hours trying Grok Build on an existing codebase, watching it silently change behavior without a warning. Not a wrong output.
A silent behavioral shift that the user only caught by accident.
And then the kicker: the terminal won't even let you copy-paste error messages.
This is the reception for xAI's big coding agent play. Grok Build landed on May 25, 2026 as an early beta. A terminal-native CLI that directly challenges Claude Code and Codex CLI.
The hype was loud. Elon posted. The xAI blog went up.
A bunch of tech outlets ran the headline. But the real story is what happened after people actually installed it.
So what is everyone arguing about?
Grok Build is not a chat interface or a VS Code plugin. It runs in your terminal. You point it at a project, describe what you want, and it plans, searches your codebase, writes code, and shows diffs for review.
The install is one line: curl -fsSL https://x.ai/cli/install.sh | bash. You authenticate with your xAI account and you're in.
The default mode is Plan Mode. Before touching a single file, Grok Build proposes a step-by-step plan. You approve it, comment on individual steps, or rewrite it entirely.
Nothing runs until you sign off. This directly fixes the thing developers hate most about coding agents.
Here's the problem it solves: an agent does something wrong, and by the time you notice, three other things have already changed downstream. Plan Mode puts a checkpoint between "task given" and "codebase modified."
Claude Code does not have this natively. That is a genuine edge.
But the architecture goes deeper. Grok Build spawns up to eight parallel subagents, each working in its own Git worktree. They do not step on each other.
An evaluation layer called Arena Mode scores competing outputs before you review. Larger refactors that would take one agent an hour can be parallelized across several agents working in isolation.
It also ships with MCP support. It picks up AGENTS.md conventions. And it supports headless mode via the -p flag for CI pipelines.
The gap nobody is glossing over
xAI published a score of 70.8% on SWE-Bench Verified for the model powering Grok Build. Claude Opus 4.7 Adaptive sits at 87.6%. OpenAI's Codex is at 85%.
That is a 17-point gap. Not a rounding error. Not a methodology disagreement.
xAI's response is that benchmarks don't fully reflect real-world engineering. That is technically true of every benchmark. It does not close a 17-point gap.
On simple, scoped tasks the difference may be invisible. On complex multi-file work, it shows up as more failed attempts, more reverted diffs, and more human review time.
One reviewer reported hallucinated edits under heavy load. Corrupted Dockerfiles from ambiguous prompts.
The competitive context makes it worse. Codex has passed three million weekly active users. Claude Code has driven Anthropic to $30 billion in annual recurring revenue.
Grok Build enters with none of that production history.
And then there's the pricing
The pricing structure is where things get interesting. Access started at \(300 per month for SuperGrok Heavy. Then they opened it to SuperGrok at \)30 and X Premium Plus at \(40, with a promotional tier at \)99.
Claude Code costs $20 flat.
On HN, the sentiment was blunt. "Only \(300 a month. (Or \)3,000 a year.) The xAI casino wants all your money even if you don't use it for a month."
Another user: "I'm not spending $300 a month on something my employer will never approve me to use."
The API pricing is more reasonable. \(1 per million input tokens and \)2 per million output tokens. And a Grok Build 0.1 model is available through OpenRouter and Vercel AI Gateway.
But the subscription model for the CLI itself is polarizing. At \(30, it's competitive. At \)300, it's a tough sell against Claude Code at $20.
The thing that actually surprised me
The TUI is genuinely impressive. I expected something slapped together. But one of xAI's engineers confirmed on HN that the terminal interface is written in Rust using Ratatui.
Proper vim keybindings. Mouse support. Careful alt-screen rendering.
They put real work into making the terminal experience feel polished.
And the binary is local-first. Your source code, credentials, and project data stay on your machine. They don't get transmitted to xAI's servers for every operation.
For a company that could easily justify cloud-only processing, that's a meaningful choice.
But here's the detail that made me pause: someone pointed out you can't copy-paste errors from the Grok Build terminal. A basic UX gap in a tool that wants to replace your existing coding workflow.
Small things like this are why "v0.1" matters.
What I actually think about coding agents in 2026
i spent an afternoon digging through HN threads, Reddit posts, and review sites to understand the reception. And what i found is that the conversation around Grok Build is less about Grok Build itself and more about where we are with coding agents in general.
Everyone is tired of benchmarks. Everyone is tired of announcements. What people actually want is a tool that doesn't silently break their codebase.
And Grok Build, for all its architectural ambition, is a v0.1 that needs the model to catch up to the interface.
The architecture is ahead of the model. That's the honest take.
If you're on SuperGrok already, try it. It costs nothing extra. But if you're deciding between Claude Code at \(20 and Grok Build at \)30, the benchmark gap is real and you will feel it on complex tasks.
Speaking of beautiful terminal tools
This whole research detour reminded me of the time i spent a weekend building a TUI for a side project using Bubble Tea in Go. Spent two days getting the layout right, another day on keybindings, and on Monday realized nobody would ever use it.
i built it for a problem nobody had.
That's kind of where Grok Build is right now. Beautiful terminal, real architectural thinking, and a model that needs maybe one more training run to justify the hype.
But at least Ratatui produces gorgeous terminal UIs. i could say the same about my Bubble Tea experiment. Except nobody used it.
We'll see if people actually use Grok Build.
Who should actually care about this
Let me be direct.
Individual dev paying out of pocket? Grok Build at \(30 is worth trying. At \)300, it is not.
The model quality is not there yet. You will get more value from Claude Code at $20 for the same tasks.
Team on xAI infrastructure or with SuperGrok enterprise access? The parallel subagent architecture is genuinely interesting. Plan Mode is a real differentiator.
But you need to accept that the underlying model will produce more bad outputs than Claude or Codex. Your team will spend more time reviewing.
Building agent orchestration tools? The ACP support matters. The headless mode and API access mean you can build custom workflows on top of Grok Build in a way that is harder with Claude Code's more limited automation surface.
For everyone else: wait for the next model iteration. The architecture is promising.
The model needs to catch up. That is the honest assessment.
One last thing
i keep thinking about that Reddit comment. Three hours watching a tool silently change code behavior. The user caught it by accident.
Not because the tool warned them. But because they happened to look at the diff before committing.
That is the real problem Grok Build needs to solve. Not the SWE-Bench score. Not the pricing.
The trust gap.
A terminal agent that quietly makes things worse while looking like it's helping is worse than no agent at all.
Plan Mode is a step in the right direction. Parallel agents are cool.
But trust is earned one clean diff at a time.
And right now, Grok Build is still earning it.