Grok Build vs Claude Code (and Codex): Which to Pick in 2026

If you've spent the past year building a Claude Code setup (a CLAUDE.md you trust, MCP servers you'd hate to reconfigure), Grok Build is awkwardly tempting. It reads all of that with zero configuration, it's open source, and it finishes a typical task for about two-thirds of Claude Code's cost.

For this Grok Build vs Claude Code comparison, I went through both tools' documentation, the current Artificial Analysis coding agent results, and the published hands-on tests, and I checked every price, model, and spec below against the vendors' own pages as of late September 2026.

That last step caught a lot, because both tools swapped their default model in the same week: Grok 4.7 landed on September 21 and Claude Opus 5.5 on September 22.

The right pick comes down to the work your agent does all day and what you already pay for, and OpenAI's Codex is a strong third option if you're on ChatGPT. (If you haven't narrowed it to these two yet, the wider field of AI coding agents is worth a look first.)

Grok Build vs Claude Code at a glance

Claude Code is still the stronger agent on hard, messy work. With Sonnet 5.5, it scores 68 to Grok Build's 56 on the Artificial Analysis Coding Agent Index as of September 29, 2026 (66 on the default Opus 5.5), and its lead comes mostly from tasks that live in the terminal.

Grok Build is cheaper per task, faster to finish, and open source, and it reads Claude Code's config files so well that trying it costs almost nothing.

If you pay for Claude Pro or Max, keep Claude Code as your main agent and give Grok Build a side job. If you're starting fresh, Grok Build is free to try and deserves a real trial, and Codex is the cheapest per task if you already pay for ChatGPT.

Prices are US, billed monthly, as of September 2026. Index scores and cost per task come from Artificial Analysis's Coding Agent Index v1.5.

What is Grok Build?

Grok Build is SpaceXAI's terminal coding agent (xAI now brands itself SpaceXAI across its site), a full-screen command-line tool that reads your repo, plans changes, edits files, and runs commands, the same job Claude Code does.

It launched in early beta on May 25, 2026, reached version 1.0 on August 7, and has run Grok 4.7 as its default model since September 21.

The pitch leans hard on compatibility. Grok Build reads AGENTS.md and CLAUDE.md, picks up your existing Claude Code skills, plugins, hooks, and MCP servers, and even accepts Claude Code's flag names (so muscle memory like --dangerously-skip-permissions keeps working, for better or worse).

It's open source under Apache-2.0, written in Rust, and runs on macOS, Linux, WSL, and Windows. You can use it interactively, headlessly with grok -p for scripts and CI, or inside editors that speak the Agent Client Protocol (ACP).

Grok Build is now free to try, according to SpaceXAI's product page, after launching for SuperGrok and X Premium+ subscribers. Paid SuperGrok plans run from $10 a month (Lite) to $300 (Heavy), but $30 SuperGrok is the first tier whose plan card lists "More powerful coding tools."

Grok Build, grok-build, and the other Grok Build

The naming is a mess, so it's worth untangling before the comparison:

• Grok Build (the CLI): the terminal coding agent compared here.
• grok-build-0.1: an older coding model on SpaceXAI's API, also called grok-code-fast-1, with a 256K context window. If you've seen Grok Build described with a 256K limit, that number belongs to this model, since Grok 4.7 has 500K.
• Grok Build in the Grok app: a separate web and mobile app builder that happens to share the name. It's been on every Grok plan since August 19, 2026.

What is Claude Code?

Claude Code is Anthropic's agentic coding tool, and it set the template Grok Build now copies, right down to the config file names. It runs in your terminal, VS Code, JetBrains, a desktop app, the browser, iOS, Android, and Slack, a much wider footprint than Grok Build's terminal-plus-ACP approach.

Since version 2.1.280, every paid plan defaults to Claude Opus 5.5, released September 22, 2026, with a 1M-token context window on every plan, Pro included. You can switch to Sonnet 5.5 or Anthropic's top model, Fable 5.1, whenever a task calls for it (on Pro, Fable bills to usage credits).

Claude Code comes with every paid Claude plan (Free doesn't include it) and shares that plan's usage with regular Claude chat. It isn't open source, either: the license on its GitHub repo reads "All rights reserved."

Team seats include it too, so if your company is already weighing Claude against alternatives to Microsoft 365 Copilot for everyone, your developers' coding agent comes bundled.

1. Pricing and usage limits

On subscription price, Claude Pro undercuts SuperGrok, the first Grok plan that lists coding tools: Claude Pro costs $20 a month billed monthly, against $30 for SuperGrok.

The tiers above scale similarly: Max 5x is $100 and Max 20x is $200 on the Claude side, while SuperGrok Plus is $100 and Heavy is $300.

Per token, the gap flips. Grok 4.7 costs $2 per million input tokens and $6 per million output tokens on SpaceXAI's API (doubling once a prompt passes 200K tokens), while Opus 5.5 costs $4 and $20. That's under a third of Anthropic's output price, which adds up fast on long agent runs.

Artificial Analysis puts Grok Build with Grok 4.7 at $8.82 per task, against $13.04 for Claude Code with Opus 5.5 and $14.19 with Sonnet 5.5, a closer gap than the token prices suggest.

Grok 4.7 is chatty (about 81,000 output tokens per task on Artificial Analysis's intelligence tests), so its cheaper rate doesn't carry over one-for-one.

Sonnet 5.5 is the odd one here. It's half the per-token price of Opus 5.5 ($2 and $10), yet it costs more per task, because Claude Code with Sonnet ran about 266 turns and 27.7 million tokens per task in Artificial Analysis's runs.

How each one meters your usage

Claude's plans use two limits at once: a session allowance that resets every five hours and a weekly cap across all models, with your reset time listed under Settings > Usage. Max 5x and Max 20x give you five or 20 times Pro's per-session allowance.

SpaceXAI doesn't publish per-plan Grok Build numbers. Paid plans share one weekly usage pool across Grok products, Build included, and a higher plan gets a larger weekly allowance.

If you'd rather skip subscriptions, both tools also run on API keys. Grok Build reads XAI_API_KEY, which is also how you run it in CI, and Claude Code moves to API billing when you sign in with a Claude Console account.

2. Agent architecture: plan mode, subagents, and worktrees

Both tools will plan before they touch anything if you ask. In Grok Build's plan mode, the agent drafts a plan, and its edit tools stay blocked until you approve.

Claude Code's plan mode works the same way, and its opusplan setting uses Opus to plan and Sonnet to carry the plan out, which spends the expensive model only where it counts.

One caveat from SpaceXAI's own docs: plan mode blocks the edit tools, but bash still runs under your normal permissions and can write files through redirection, and subagents aren't held by plan mode at all. Keep that in mind before you point it at a repo you care about.

Subagents look almost identical on paper. Grok Build's run in parallel as independent child sessions with their own context. It ships general-purpose, explore, and plan types, lets you define your own in .grok/agents/, and can give each subagent its own git worktree so parallel edits don't collide.

Grok Build also has /create-workflow, which saves reusable workflows that fan work out to subagents and verify the results in the background.

Claude Code has the same building blocks: subagents with their own context window, tools, and permissions, plus worktree isolation for parallel sessions. Its agent teams feature, where several Claude sessions coordinate on one job, is still experimental and switched off by default.

The difference shows up in how long each one sticks with a problem. On Artificial Analysis's runs, Claude Code with Sonnet 5.5 took about 266 turns and 1.5 hours per task, while Grok Build took about 163 turns and 39 minutes.

Grok Build's speed is great for throughput and less great when a problem needs the extra grinding.

3. Memory and instruction files

This is where Grok Build made its smartest move. It reads AGENTS.md, CLAUDE.md, and CLAUDE.local.md, plus rule folders from .grok/, .claude/, and .cursor/, so a repo set up for Claude Code (or Cursor) works in Grok Build on day one.

It also loads MCP configs from ~/.claude.json and .mcp.json, runs the hooks in your .claude/settings.json, and can import your past Claude Code sessions with grok import.

The compatibility only runs one way. Claude Code reads CLAUDE.md plus its own auto memory, and it falls back to AGENTS.md when there's no CLAUDE.md, but it won't read anything in .grok/. If you start in Grok Build and later move back, your Grok-specific agents and workflows stay behind.

If you plan to run both, keep your shared instructions in AGENTS.md and have CLAUDE.md import it. By default, Claude Code reads only CLAUDE.md when both files exist, so the import keeps the two agents working from the same rules.

4. Grok Build vs Claude Code benchmarks and models

Grok Build runs Grok 4.7 by default, with a 500,000-token context window, a May 2026 knowledge cutoff, and four reasoning effort levels up to xhigh, its highest setting.

There's also Grok 4.7 Fast, the same model on faster hardware at twice the token rate, which is available only in Grok Build and Cursor.

Claude Code defaults to Opus 5.5 with a 1M-token window. You can switch to Sonnet 5.5 ($2 and $10 per million tokens) for cheaper work or Fable 5.1 ($10 and $50) when you want Anthropic's strongest model.

Grok Build is also the more open of the two about models. You can point it at any OpenAI-compatible or Anthropic Messages endpoint in ~/.grok/config.toml, so you can even run Claude inside it. Claude Code sticks to Claude models, and Anthropic doesn't support routing it to non-Claude models through any gateway.

What the benchmarks measure

Artificial Analysis tests each agent inside its own harness and averages three benchmarks into its Coding Agent Index v1.5:

• DeepSWE v1.1: 113 long-horizon engineering tasks that each require changing an existing repository.
• Terminal-Bench 4.0: 66 terminal tasks spanning software engineering, machine learning, scientific computing, security, and system administration.
• SWE-Atlas-QnA: 124 questions about a repository that the agent answers by tracing the code and explaining how it behaves.

Here's how the two compare as of September 29, 2026, using each tool's best-scoring setup:

Claude Code with Opus 5.5, the default most people will run, scores 66 at about $13 a task, so on this index Sonnet 5.5 edges its bigger sibling, even though it costs more per task.

Grok Build edges Claude Code on DeepSWE, the long repo-change tasks, and trails by 33 points on Terminal-Bench, which the average hides.

If your work is mostly "change this code in this repo," the gap is smaller than the headline suggests. If your agent spends its day running builds, fixing environments, and debugging from the shell, Claude Code's lead is the one to weigh.

The vendors' own launch numbers point the same way: SpaceXAI reports 37.6% on Terminal-Bench 4.0 for Grok 4.7 at xhigh, and Anthropic reports 66.4% for Opus 5.5. Vendors test at different effort settings, so lean on Artificial Analysis for cross-vendor comparisons.

5. Security and privacy

Grok Build had a rough July. A researcher publishing as cereblab found that version 0.2.93 uploaded whole Git repositories, commit history included, to a SpaceXAI storage bucket, even files it never opened. On one test repo, the model traffic came to about 192 KB while the upload moved 5.10 GiB.

Turning off the "Improve the model" setting didn't stop it, and a tracked .env file the agent read went out unredacted. SpaceXAI switched the uploads off server-side on July 13, and Elon Musk promised the uploaded data would be deleted.

If you used Grok Build before mid-July on a repo with secrets anywhere in its history, rotate those credentials (a boring afternoon, and cheaper than finding out the hard way). The CLI's /privacy command shows and toggles data retention.

Day to day, Claude Code sandboxes Bash at the operating-system level (Seatbelt on macOS, bubblewrap on Linux and WSL2, with no support on native Windows), and its auto permission mode uses a classifier to decide what needs your sign-off.

Grok Build asks before anything you haven't already allowed, which is a sensible default. Its sandbox, though, is off until you turn it on, and the sandbox's network blocking only works on Linux.

Hooks are the other difference worth knowing. In Grok Build, PreToolUse is the only hook that can block an action, and a hook that times out or crashes lets the tool call go through, so don't make Grok Build hooks your only safety net.

6. Extensibility: MCP, skills, hooks, and plugins

Feature for feature, the two are closer here than anywhere else. Both support MCP servers, SKILL.md skills that show up as slash commands, hooks, plugins, and headless runs for CI, and Grok Build loads Claude Code's plugins and marketplaces as they are.

Claude Code's hooks go further: they can be shell commands, HTTP endpoints, MCP tool calls, LLM prompts, or subagents, while Grok Build's are shell or HTTP. Claude Code also has the Agent SDK for Python and TypeScript and an official GitHub Actions integration.

Grok Build's extras are smaller but handy: a marketplace tab inside the terminal UI, background tasks you can monitor, and /loop prompts that rerun on a schedule (every 60 seconds at most often, and they expire after seven days, which saves you from a forgotten loop billing you all month).

7. Where each one runs

Grok Build lives in the terminal. You get the interactive full-screen interface, a headless mode with plain, JSON, or streaming JSON output, and ACP for editors that support it, but there's no first-party IDE extension.

Claude Code goes wherever you are: the terminal, VS Code, JetBrains, a desktop app, the browser, iOS and Android, Slack, and GitHub. If you like starting a task from your phone and reviewing the diff at your desk, Grok Build can't do that yet (Codex can, on iOS).

Same task, both tools: what the hands-on tests found

Aashi Dutt ran Grok Build (on Grok 4.6) and Claude Code (on Opus 5) through the same rigged churn-prediction task for DataCamp in August 2026. Both models have been replaced since, so read it as a character study of the two agents.

The dataset had three planted problems: a column that leaked the answer, heavy class imbalance, and customers who showed up on both sides of the train and test split. The four rounds went like this:

Dutt's verdict was that Grok Build "requires less verification," while Claude Code gave more and needed checking, including a count of repeat customers it got wrong. That fits the benchmark picture, where Grok Build wraps up quickly, and Claude Code keeps working the problem.

Claude Code's extra digging did turn up a production bug in Dutt's own code. It's also one conversation per agent, which is a small sample.

What developers are saying about Grok Build

Daniel Lemire, a computer science professor known for performance work, wrote about Grok Build in June 2026. It performed "superbly" on a dashboard for his school's statistical data, and an ambitious C++ game remake stumbled early on repeated edits to one file before going well.

Robert Mill switched to Grok Build in July after repeatedly hitting Anthropic's usage limits, and the main praise was speed: a dev server running in under two minutes and plans that came back in seconds. Both reports predate Grok 4.7.

Can you use Grok Build and Claude Code together?

Yes, and for a lot of developers it's the best setup. SpaceXAI publishes an official Claude Code plugin, grok-build-plugin-cc, that lets Claude Code hand reviews, rescue tasks, and whole sessions over to the Grok Build CLI.

The Grok side only needs the CLI installed and signed in, since Grok Build already reads your CLAUDE.md, skills, and MCP servers. A reasonable split is Claude Code for the hard, long-running changes and Grok Build for second-opinion reviews, parallel grunt work, and cost-sensitive jobs.

What you can't do is run Grok 4.7 inside Claude Code: Anthropic doesn't support routing Claude Code to non-Claude models, and SpaceXAI has deprecated its Anthropic-compatible endpoint. Running Claude inside Grok Build works fine through its Anthropic Messages backend.

What about Codex? Grok Build vs Codex

OpenAI's Codex is the third terminal agent in this conversation, and if you pay for ChatGPT, you already have it. Codex comes with ChatGPT Plus ($20 a month), Pro (from $100), and Business ($25 per user monthly) across the web, CLI, IDE extension, and iOS.

Free and Go users are getting GPT-6 Luna in the desktop app as it rolls out, and OpenAI recommends GPT-6.1 Sol for complex coding, at $2 per million input tokens and $10 per million output tokens on the API, with a context window just over 1M.

GPT-5.5 leaves Codex on October 14, 2026, so if your setup pins it, now's the time to move.

On the Artificial Analysis Coding Agent Index, Codex with GPT-6.1 Sol scores 63 at $1.04 per task, seven points ahead of Grok Build's 56 at about an eighth of the cost. Claude Code with Sonnet 5.5 still tops the chart at 68.

On Terminal-Bench 4.0 alone, Codex with GPT-6.1 Sol hits about 55%, roughly 21 points ahead of Grok Build.

Codex reads AGENTS.md, its CLI is open source under Apache-2.0 (the IDE extension isn't), and Codex cloud runs tasks in isolated cloud environments that you can start from GitHub, GitLab (in beta), Linear, or Slack.

So Grok Build vs Codex mostly comes down to what you already pay for. Codex with GPT-6.1 Sol is the cheapest per task of the three by a wide margin, and for ChatGPT subscribers it costs nothing extra to try.

Grok Build is the pick if you want Claude Code compatibility or a fully open-source agent. And if you're still torn between Claude and ChatGPT as subscriptions, the bundled coding agent is a fair tiebreaker.

‍

{{templates}}

‍

Which should you pick?

Choose Claude Code if you:

• Spend your day in the shell: its 33-point Terminal-Bench lead pays off when the agent has to run builds and fix environments.
• Already pay for Claude Pro, Max, or Team: Claude Code is included, with Opus 5.5 and a 1M-token window on every paid plan.
• Want it everywhere: IDEs, desktop, browser, phone, and Slack are all first-party.
• Want OS-level sandboxing for Bash: on macOS, Linux, and WSL2.

Choose Grok Build if you:

• Want a cheaper second agent: $8.82 per task on Artificial Analysis's runs, with your Claude Code setup working as is.
• Want an agent you can read and fork: the whole CLI is Apache-2.0 on GitHub.
• Run lots of parallel, well-scoped tasks: it finishes faster and in fewer turns.
• Want to bring your own models: including Claude, through one CLI.

Choose Codex if you:

• Already pay for ChatGPT: Codex is included on every ChatGPT plan, with the web, CLI, IDE, and iOS versions from Plus up.
• Want the lowest cost per task: $1.04 with GPT-6.1 Sol.
• Want cloud tasks: started from GitHub, GitLab (in beta), Linear, or Slack.

Hold off on Grok Build if you:

• Work on sensitive repos under strict review: the July upload fix was a server-side switch, so clear it with your security team before a wider rollout.
• Live in your IDE: there's no first-party extension, only ACP.

Give both agents the same ticket this week

Grok Build is compatible enough with Claude Code that a real test is cheap. Install it, open a repo you already use with Claude Code, and hand both agents the same real ticket on separate branches. Then compare the diffs, how often you had to step in, and what each run cost you.

My read on Grok Build vs Claude Code, after all the numbers, is that Claude Code stays the better main agent for hard work, Grok Build earns a spot as the cheaper second agent, and Codex is the easy win for anyone already on ChatGPT.

Whichever one you keep, coding is only part of shipping, and the on-call bug reports, changelogs, and stale pull requests still pile up in Slack. That's the kind of work Lindy, an AI teammate that lives in Slack, can pick up for product and engineering teams.

‍

{{cta}}

‍

FAQ

Does Grok have a coding agent like Claude Code?

Yes, Grok Build is SpaceXAI's terminal coding agent, and it works much like Claude Code: it reads your repo, plans changes, runs commands, and edits files. It launched in May 2026, runs Grok 4.7 by default, and reads Claude Code's CLAUDE.md files, skills, and MCP servers with zero configuration.

Is there anything better than Claude Code for coding?

No coding agent beats Claude Code on the Artificial Analysis Coding Agent Index right now: it leads at 68 with Sonnet 5.5 as of September 2026. Codex with GPT-6.1 Sol (63) and Grok Build (56) cost less per task, and the wider set of Claude alternatives covers writing and business work too.

Can I connect Grok to Claude Code?

Yes, through SpaceXAI's official grok-build-plugin-cc plugin, which lets Claude Code hand reviews, rescue tasks, and whole sessions to the Grok Build CLI. Running Grok models inside Claude Code itself isn't supported, because Anthropic doesn't support routing it to non-Claude models and SpaceXAI deprecated its Anthropic-compatible endpoint.

How good is Grok Build?

Grok Build is a strong, cheaper second agent: it scores 56 on the Artificial Analysis Coding Agent Index with Grok 4.7 at about $8.82 per task, edging Claude Code on long repo changes and trailing badly on Terminal-Bench 4.0. Review its July repo-upload incident before using it on sensitive code.

Does Grok Build run locally?

Grok Build runs on your machine as a local CLI on macOS, Linux, WSL, and Windows, but its default model runs in SpaceXAI's cloud, so your prompts and the files it reads leave your machine. You can also point it at your own model endpoint in ~/.grok/config.toml.