Grok Build vs Claude Code Comparison (2026): 32 Subagents vs 1M Context

Quick Verdict

Grok Build vs Claude Code by Version, Release Date, and Price (September 2026)

Both agents ship several times a week, which is why comparisons written in spring 2026 describe products that no longer exist. The table records what each tool was at when this page was last checked, so you can measure how far it has moved since.

Two things fell out of the version check. Grok Build has no git tags and no GitHub releases; the only version marker is the xai-grok-version crate, whose description reads "Lockstepped grok CLI version" and which sat at 1.0.24 on the September 9, 2026 sync commit. And xAI ships a second, cheaper coding model, grok-build-0.1, at $1 in and $2 out per million tokens on a 256K context. It is priced in the public model table but is not the Grok Build default.

Architecture: Parallel Breadth vs Reasoning Depth

The fundamental difference between these two tools is architectural, and it shapes everything else.

Grok Build: Plan, Then Fan Out

Grok Build cycles three session modes with Shift+Tab: plan, auto, and always-approve. Plan mode gates file edits so only the session plan file can be written until you approve it, and that gate is independent of the permission mode. You enter it with /plan and reopen a plan with /view-plan.

Once the plan is approved, the main agent delegates to subagents. There are three built-in types: general-purpose with full capability, explore which can read, list and search but cannot run a shell or edit, and plan which drafts an implementation plan with the same restrictions. You add or override types under .grok/agents/.

The tradeoff is context. Each subagent is an independent child session with its own context that returns a summary to the parent when it finishes. For a task where the relationship between distant files matters, no single agent sees the whole picture.

Claude Code: One Agent, Deep Context

Claude Code takes the opposite default. One agent on a 1M token context window that can hold most repositories at once. Instead of fanning out first, it reasons within a single context, tracking dependencies across files and producing changes that are internally consistent.

Claude Code can spawn subagents, but this is opt-in rather than default. The primary workflow is sequential: understand the full picture, plan the change, execute across files. This produces more deterministic output and takes longer per task.

Multi-Agent Approach Comparison

Both tools support multi-agent workflows. The default behavior, the concurrency ceiling, and how you raise it differ.

Grok Build: 32 Concurrent Subagents, Not 8

Launch coverage in May 2026 put Grok Build at 8 parallel agents. The shipped source says otherwise. DEFAULT_MAX_CONCURRENT is 32 in the subagent admission module at crates/codegen/xai-grok-tools/src/implementations/grok_build/task/admission.rs, whose file comment reads "Session-scoped subagent spawn limits, enforced by the coordinator."

Two environment variables control it, and neither appears in the published settings reference at docs.x.ai. GROK_MAX_CONCURRENT_SUBAGENTS sets the cap, and the source notes a limit "can be adjusted but never disabled": a value of 0 is clamped to 1. GROK_SUBAGENT_LIMIT_BEHAVIOR chooses what happens to a spawn that arrives at the ceiling, accepting queue or fail. The default is queue, so a 33rd subagent waits rather than erroring, and any other value logs a warning and falls back to queue.

Claude Code: Subagents On Demand

Claude Code's subagents are spawned explicitly when parallelism is needed. The primary agent coordinates, delegates specific investigation or implementation tasks, and synthesizes results. Each subagent gets its own context window and tool access, and SubagentStart and SubagentStop hooks fire around each one.

The key difference: Claude Code's subagents are coordinated by a primary agent that holds the full context in a 1M token window. Grok Build's subagents are more autonomous, each returning a summary to a parent working within a 500K window.

Arena Mode, the automatic best-of-N evaluator that May 2026 coverage described, is gone from the documentation. The string "arena" does not appear anywhere in docs.x.ai, including its llms.txt index, and no /arena command exists in the Grok Build command table as of September 14, 2026. xAI has not published a note explaining the removal. The nearest current equivalent is manual: spawn several subagents on the same task and compare their summaries yourself.

Pricing Comparison

Token pricing is verifiable from both vendors. Subscription pricing is only verifiable on the Anthropic side: xAI's consumer pricing pages at x.ai and grok.com returned 403 and 404 respectively when checked on September 14, 2026, so the $299 per month SuperGrok Heavy figure this page carried from the May 2026 beta is marked unverified below rather than repeated as fact.

Cost Per Successful Task

Raw subscription price is only half the equation, and the multi-agent side of it cuts against Grok Build. Thirty-two concurrent subagents each carry their own context, so the token cost of a fanned-out task scales with the number of children, not with the size of the change. Claude Code's single-agent default spends more per turn on a bigger context and fewer turns overall.

The weekly pool is the practical constraint on the Grok side. A developer who built an open-source macOS-binary compatibility layer with Grok Build reported on Hacker News in August 2026 that despite using Grok 4.5 at medium effort with "no sub-agents, just standard chats and Plan Mode," they "hit the weekly usage limits all the time and had to sit around waiting for the cooldowns." That was a single-agent workload. Fanning out to 32 subagents draws on the same pool.

Grok Build vs Claude Code vs Codex

The three-way question comes up because all three are terminal agents from frontier labs, and two of the three are open source. Prices below are from each vendor's own pricing page on September 14, 2026.

Codex is the cheapest way in. OpenAI sells a Go tier at $8 per month for lightweight coding, below the $20 Plus tier that includes web, CLI, IDE and iOS access, and its Pro tier starts at $100 per month for 5x or 20x higher rate limits than Plus. Read the full head-to-head on Codex vs Claude Code.

The Terminal-Bench 4.0 and CursorBench 3.2 numbers in that table for Codex are Anthropic's measurements of GPT-5.6 Sol, published alongside its own scores, not OpenAI's. xAI has not published a Terminal-Bench 4.0 number at all; its Grok 4.6 announcement reports Terminal-Bench v3.0 at 26%, up from 15.7% for Grok 4.5, which is a different harness version and does not compare.

Benchmarks

Benchmark data checked September 14, 2026. The comparison is harder than it was in May, because the two labs now publish on almost disjoint benchmark sets.

For context, other terminal agents: OpenAI Codex runs GPT-5.6 Sol, which Anthropic measured at 67.2% on CursorBench 3.2 and 37.3% on Terminal-Bench 4.0. Gemini CLI now ships as Antigravity CLI on Google's shared Antigravity harness.

Context Window and Codebase Handling

Context management is where the architectural difference becomes most visible in daily use.

Claude Code: 1M Tokens, Single Agent

Claude Opus 5, Claude Sonnet 5 and Claude Fable 5.1 all carry 1M token context windows with 128K max output. On the current tokenizer, Anthropic puts 1M tokens at roughly 555,000 words. That is large enough to hold most repositories in a single context, so the agent reads the full structure, understands architectural patterns, and tracks dependencies across files. Long sessions still benefit from proactive compaction, and PreCompact and PostCompact hooks fire around it.

Grok Build: 500K Per Agent, Distributed

Grok Build's default model, grok-4.6, carries a 500K context window. Crossing 200K prompt tokens doubles the price to $4 in and $12 out per million. Each subagent is an independent child session with its own context that returns only a summary to the parent, so aggregate capacity across 32 children is large but no single agent sees the whole picture.

For large codebases with deeply interconnected modules, this is a meaningful tradeoff. Refactoring an auth module that touches billing, API routes, and database schemas works better when a single agent holds all four concerns at once. For feature additions confined to a directory, the distributed approach has less downside.

Using Grok in Claude Code

The mechanism exists, and xAI has marked it for removal. Both halves matter, so here is the exact state as of September 14, 2026.

What xAI exposes

xAI serves a POST /v1/messages endpoint at https://api.x.ai/v1, documented as "compatible with the Anthropic API." Claude Code reads ANTHROPIC_BASE_URL to override the API endpoint and ANTHROPIC_AUTH_TOKEN as the value of the Authorization header, prefixed with Bearer. Wiring those two together is the whole configuration:

export ANTHROPIC_BASE_URL=https://api.x.ai/v1
export ANTHROPIC_AUTH_TOKEN=<your xAI API key>
export ANTHROPIC_MODEL=grok-4.6
claude

Why it is not a setup to build on

xAI's own REST reference files that endpoint under "Legacy & Deprecated" and prints, three times on the page, "Deprecated: The Anthropic SDK compatibility is fully deprecated. Please migrate to the Responses API or gRPC." The /v1/complete endpoint carries the same notice. xAI has not published a shutdown date, so this is a documented deprecation without a documented end.

Claude Code also degrades on a non-first-party host, and its environment variable reference says so explicitly. Pointing ANTHROPIC_BASE_URL at a non-first-party host disables MCP tool search by default, recoverable with ENABLE_TOOL_SEARCH=true only if the endpoint forwards tool_reference blocks. As of v2.1.196, Remote Control is disabled whenever ANTHROPIC_BASE_URL points anywhere other than api.anthropic.com. Anthropic documents third-party routing for Amazon Bedrock, Google Cloud, Microsoft Foundry and LLM gateways, all serving Anthropic models. It does not document running another lab's model in Claude Code.

Ecosystem: MCP, Hooks, Plugins

Both tools support the same categories of extensibility. The hook surface is where the gap is measurable rather than impressionistic.

Grok Build fires 14 lifecycle events: SessionStart, SessionEnd, UserPromptSubmit, PreToolUse, PostToolUse, PostToolUseFailure, PermissionDenied, Stop, StopFailure, Notification, SubagentStart, SubagentStop, PreCompact and PostCompact. Only PreToolUse can block. Everything else is fail-open: a timeout, crash or malformed output is recorded in the session and the tool call proceeds anyway.

Claude Code documents 33 as of v2.1.271, a superset of Grok's plus Setup, InstructionsLoaded, UserPromptExpansion, MessageDisplay, PermissionRequest, PostToolBatch, TaskCreated, TaskCompleted, TeammateIdle, ConfigChange, CwdChanged, DirectoryAdded, FileChanged, WorktreeCreate, WorktreeRemove, PreModelSwitch, PostModelSwitch, Elicitation and ElicitationResult. The hooks system also supports prompt-based and agent-based handlers, not just shell commands and HTTP endpoints.

When to Use Which

Choose Grok Build When

Choose Claude Code When

Frequently Asked Questions

What is Grok Build?

Grok Build is xAI's terminal coding agent. It shipped in early beta on May 14, 2026, opened to SuperGrok and X Premium+ subscribers on May 25, 2026, and became open source on July 14, 2026 at github.com/xai-org/grok-build under Apache-2.0. The CLI is at version 1.0.24 as of the September 9, 2026 monorepo sync. It runs grok-4.6 by default, cycles plan, auto and always-approve modes with Shift+Tab, spawns general-purpose, explore and plan subagents, and supports MCP servers, hooks, skills, plugins, worktrees, headless mode via -p, and ACP.

How does Grok Build's Arena Mode work?

It does not, any more. Arena Mode is no longer a documented Grok Build feature. May 2026 launch coverage described an automatic best-of-N evaluator that scored competing agent outputs. As of September 14, 2026 the string "arena" does not appear anywhere in the documentation at docs.x.ai, including its llms.txt index, and no /arena command exists in the command table. xAI has not published a note explaining the removal. The current equivalent is manual: spawn several subagents on the same task and compare their summaries yourself.

How much does Grok Build cost compared to Claude Code?

On the API, grok-4.6 costs $2 per million input tokens and $6 per million output tokens below a 200K prompt, with cached input at $0.50, doubling to $4 and $12 above 200K. Claude Opus 5 costs $5 in and $25 out, Claude Sonnet 5 $2 and $10, Claude Fable 5.1 $10 and $50. On subscriptions, Claude Code is included in Claude Pro at $20 per month, $17 billed annually, and in Claude Max from $100 per month. Grok Build signs in with a SuperGrok or X Premium+ subscription, but xAI's consumer pricing pages returned 403 and 404 when checked on September 14, 2026, so the widely repeated $299 per month SuperGrok Heavy figure could not be verified.

Which has better benchmark scores?

CursorBench 3.2 is the only benchmark both vendors publish. Anthropic reports 73.4% for Claude Fable 5.1 at max effort, 70.5% for Fable 5, 70.0% for Claude Opus 5 and 67.2% for GPT-5.6 Sol. xAI reports 69.9% for Grok 4.6. Anthropic has not published a SWE-bench Verified score for Opus 5; its last published figure was 80.84% for Claude Opus 4.6 in February 2026, which is why comparisons still quoting 80.8% as a current Claude Code score are stale. Terminal-Bench does not compare: Anthropic reports version 4.0, xAI reports version 3.0.

Does Grok Build support MCP servers and hooks?

Yes. Grok Build supports MCP servers, 14 hook lifecycle events, skills, plugins, marketplaces, worktrees, a four-profile sandbox, background tasks and ACP. It also reads Claude Code's files directly: AGENTS.md, CLAUDE.md, CLAUDE.local.md, .claude/rules/ and .cursor/rules/ all load as project rules, and hooks in .claude/settings.json are read with Claude tool names such as Bash, Read and Edit mapped onto Grok's automatically. Claude Code documents 33 hook events as of v2.1.271.

How do I use Grok in Claude Code?

Set ANTHROPIC_BASE_URL to https://api.x.ai/v1 and ANTHROPIC_AUTH_TOKEN to an xAI key, and Claude Code will talk to xAI's Anthropic-compatible /v1/messages endpoint. xAI's REST reference files that endpoint under "Legacy & Deprecated" and states the Anthropic SDK compatibility is fully deprecated, directing developers to the Responses API or gRPC, with no shutdown date published. Claude Code also degrades on a non-first-party host: MCP tool search is off unless ENABLE_TOOL_SEARCH=true, and Remote Control is disabled as of v2.1.196. Anthropic documents third-party routing only for Bedrock, Google Cloud, Microsoft Foundry and LLM gateways, all serving Anthropic models.

Related Comparisons

• Claude Code vs Cursor
• Codex vs Claude Code
• Gemini CLI vs Claude Code
• Claude Code vs GitHub Copilot
• Claude Code Alternatives