Grok Build
A bet on parallelism: up to eight agents on one task; a headless mode for CI already in the base tier.
Verified: Sep 12last change: Sep 21next price check: Dec 11
Facts with sources
- Left beta on August 7 as version 1.0 under the Apache 2.0 license. Source (in Russian)
- Free access appeared at launch and isn’t locked into the pricing plans: treat it as a promotion. Source (in Russian)
- Terminal-Bench 4.0: 20.30% of tasks solved on high effort, at a run cost of $3.6K. Source (in Russian)
Strengths
- Rust harness under Apache 2.0; a third-party or local model key set in the config takes priority over the xAI account.
- Goal mode switches to plan on its own and ends with a skeptic agent: for Arcade it rejected the result, then re-checked only the delta.
- Parallel subagents: Arcade got 57 subagents in about three minutes to review files; the docs say 8 by default in isolated git worktrees.
- Lead Gen Jay's head-to-head via the Claude Code bridge: Grok finished in 12 min 53 s vs about 30 min for Codex, 45 passing tests vs 33.
Weaknesses
- The 460-line installer checks neither signature nor checksum: 164 MB of executable code runs with your user rights (Edward Grishin).
- By default it reads Claude Code and Cursor configs without asking, including files holding keys, and always runs their global hooks.
- The code carries remote-controlled settings: the server can switch on telemetry and workspace file capture wherever you set no value.
- Not usable from Russia: download, sign-in and payment run through xAI domains, so a VPN and a foreign card are required.
Example from practice
Get a repo to 'clone and run in under five minutes'
Arcade gave Grok Build a goal on a Next.js demo repo: bring it to a state where a first-time contributor can clone it and run the demo in under five minutes. The agent switched itself into plan mode, went through .gitignore, migrations and environment variables, then launched a skeptic agent to verify. The skeptic found that better-sqlite3 fails under Bun, refuted the result, the main agent added prerequisites and a troubleshooting note, and the re-check covered only the delta and passed on attempt two of six. Total: 23 minutes and roughly 150,000 tokens.
source: Arcade, 'Grok Build Spawned 57 Agents in 3 Minutes. Here's Why I Wasn't Worried.'
What reviewers say
Reviewer opinions on their own tasks and hardware, not measurements. Links go to the videos; transcripts live in docs/raw-data.
- Grok 4.6 Is INSANE for $10/Month · Gui Bibeau, Sep 3The reviewer calls Grok Build bundled with the $10 X Premium a steal on price and speed, even though his 15-minute 3D scene came out 'childish' and not shareable.
- Grok Build Spawned 57 Agents in 3 Minutes. Here's Why I Wasn't Worried. · Arcade, Sep 9Two thumbs up from the reviewer for goal mode with a verifying agent and the transparent UI; his caveat is that git only protects you inside the repo.
- Everything You Need to Know About Grok Bot (Full Grok Build) · Lead Gen Jay, Aug 26The reviewer keeps the Grok subscription for the model as a second reviewer inside Claude Code, but Grok Bot did not replace his Hermes agent.
- DeepSeek Harness, Grok Build, OpenWorker, qm. ОБЗОР 4 HARNESS. ЧТО Я НАШЕЛ В КОДЕ, ВАС УДИВИТ! · Эдвард Гришин | AI и автоматизация бизнеса, Sep 1In the reviewer's view it is the richest of the four summer harnesses, but only worth installing if you already pay for Grok, and only after disabling foreign-config reading.
Who it is for
- Engineering team outside Russia that needs a CI mode
Models in this agent
Score is from our manager-task benchmark; models without a score are not measured yet.
What changed
- Grok 4.7 released: $2 per million input and $6 per million output tokens; a fast variant is twice as fast at twice the price →
- Grok Build left beta with version 1.0 under the Apache 2.0 licence
How to brief an agent like this
The four-line brief is the Foundation module: prompt structure, LLM agents and MCP, an AI policy for the team. It starts with the free open module.