xAI Grok Build Arrives: Parallel Agents and Plan Mode Shake Up Terminal Coding

The terminal coding-agent market, long a two-player game, became a three-way fight on May 14, 2026 . xAI launched Grok Build, its first agentic command-line interface for professional software engineering, directly targeting Anthropic's Claude Code and OpenAI's Codex CLI . The tool is now available in early beta exclusively for SuperGrok Heavy subscribers at a standard price of $300 per month, with an introductory offer of $99 per month for the first six months .

Grok Build enters the arena with a clear architectural thesis: raw model performance isn't everything. Instead, xAI is betting on a novel combination of massive context, mandatory human-in-the-loop safeguards, and genuine parallel execution to win over developers working on complex, multi-file codebases .

How Access Works

Grok Build is an intentionally gated product during its early beta phase. Unlike Claude Code, which is available across Anthropic's standard paid plans, and Codex CLI, which is bundled into ChatGPT paid tiers, Grok Build requires a top-tier SuperGrok Heavy subscription .

• Subscription Tier: SuperGrok Heavy at $300/month (standard) .
• Introductory Pricing: $99/month for the first six months for new subscribers , which can also be found listed as $299/month with a $99 promotional rate .
• Alternative Access: The underlying API is also available at $0.20 per million input tokens and $1.50 per million output tokens .

Key Technical Features

Grok Build is built on the Grok 4.3 beta architecture and designed to run as a Rust-based terminal UI (TUI), usable both interactively and headlessly inside CI/CD pipelines . Here are the capabilities that define it.

Plan Mode: Execution with a Human Checkpoint

Before Grok Build writes or modifies a single file, it generates a detailed step-by-step plan and presents it for user review. Developers can approve the plan, comment on specific steps, or rewrite sections entirely. Only after an explicit sign-off does the agent begin executing changes, which appear as clean diffs . This is a stricter human-in-the-loop approach than Claude Code or Codex CLI, which generally execute tasks more autonomously without a mandatory plan-approval gate . Some reports note that Grok Build's Plan Mode generates a visual graph of sub-tasks with per-node state in a dedicated terminal UI, a richer representation than the linear text plans produced by its competitors .

Parallel Sub-Agents (Up to 8)

The most architecturally significant feature is native parallelism. Instead of one agent working sequentially, Grok Build can spawn up to eight specialized sub-agents simultaneously—for example, one searching the codebase, one writing unit tests, and another modifying database schemas—then merge the results . Each sub-agent can operate in an isolated Git worktree, a feature neither Claude Code nor Codex CLI ships natively . This design is purpose-built for large, monorepo-style codebases where parallel task execution saves meaningful time.

Context Window: 2M Tokens Claimed

Grok Build's architecture claims access to a 2-million-token context window via the underlying Grok 4.3 Heavy, 16-agent architecture . In practice, the specific agentic model grok-code-fast-1 has been documented with a 256K context window, while the dedicated grok-build-0.1 model, released on May 20, 2026, is the production model now powering the CLI . A 2M-token context window, if realized in active coding sessions, represents roughly double the 1M-token context of Claude Code and would allow developers to hold an entire medium-to-large codebase in active memory simultaneously .

Protocol Compatibility: Zero Switching Cost

Grok Build deliberately adopts Claude Code's configuration ecosystem to minimize migration friction. It supports MCP (Model Context Protocol), ACP (Agent Client Protocol), and the same Skills/AGENTS.md conventions, allowing teams to drop it into existing Claude Code environments without rewriting their agent instructions or tool configurations .

Future: Arena Mode

xAI has previewed an upcoming Arena Mode, a self-evaluation harness where Grok Build will internally test and score competing code solutions against benchmarks, effectively running a tournament between its own approaches before presenting a final result .

Competitive Positioning: Benchmarks and Real-World Utility

The Benchmark Gap

On the industry-standard SWE-Bench Verified benchmark, Grok Build's initial score is notably lower than its established rivals :

• Codex CLI (GPT-5.5): 88.7% (vendor-reported)
• Claude Code (Opus 4.7): 87.6% (vendor-reported)
• Grok Build (grok-code-fast-1): 70.8%

That 70.8% score belongs to the now-deprecated grok-code-fast-1 model, not the grok-build-0.1 model powering the updated CLI released on May 20, 2026 . The score sits roughly 17 points behind the leaders, and xAI has not yet published updated benchmark numbers for the newer model . For developers who prioritize raw code-generation accuracy, the gap is significant. However, some early coverage and analysis suggests the benchmark score may not capture Grok Build's real-world advantage in parallel orchestration tasks where architectural choices matter more than single-pass accuracy .

Ecosystem Maturity and Stability

Claude Code remains the market leader on stability, developer mindshare, and ecosystem integration, with support across IDE, GitHub, Xcode, and voice interfaces . Its safety-first design and enterprise compliance track record make it the safest choice for teams that value reliability above experimentation . Codex CLI, running on GPT-5.5, is the strongest option for organizations already invested in the OpenAI ecosystem, with recent mobile and remote-dispatch features . Grok Build, by contrast, is an early beta with novel architecture but no production track record, and its $300/month price point makes it the most expensive entry in the CLI coding-agent market .

When Grok Build Might Win

The clearest use case for Grok Build is large, parallelizable tasks in monorepo environments. The combination of a massive claimed context window and native parallel sub-agents with worktree isolation is currently unmatched by Claude Code or Codex CLI . A common shorthand among early adopters captures the trade-off: "Monorepo? Grok Build. Stability? Claude Code. OpenAI ecosystem? Codex CLI" . For teams willing to tolerate beta risk in exchange for architecture that maps directly to their parallel development workflows, Grok Build is worth testing. For everyone else, Claude Code and Codex CLI remain safer, battle-tested choices today .