Grok Build 0.1

“Cheap, with retention controls that ease the IP objection — but code still reaches xAI, and at v0.1 it's a second tool, not a Claude Code replacement.”

Strategically, Grok Build 0.1 answers the top CTO objection to coding agents — shipping proprietary source to a cloud LLM — less fully than its launch coverage claimed: xAI's security FAQ says the CLI retains code data unless you turn code-data retention off in /privacy or /settings or enable team-wide Zero Data Retention, so code still reaches xAI. With those controls set it is usable for IP-sensitive work, but not for codebases that legally can't go to the cloud. The 70.8% SWE-bench Verified is competitive-not-leading, and v0.1 maturity means fewer integrations and rough edges. The sensible posture is "second tool in the belt": use it for cost-sensitive, well-scoped work, keep Claude Code as primary elsewhere. Vendor risk is the usual xAI profile (thin disclosure, SOC 2 as the only stated certification); lock-in is low via OpenAI-compatible API. The X-ecosystem angle does not meaningfully extend to coding.

“xAI is now competing across the full developer-tools surface — but its data-sovereignty wedge is thinner than launch coverage claimed; price is the real one.”

In dev-tools market terms, Grok Build signals xAI competing beyond chat into the coding-agent category Anthropic has led for a year. The claimed differentiation was local-first data sovereignty, but xAI's security FAQ says the CLI retains code data unless you turn code-data retention off in /privacy or /settings or enable team-wide Zero Data Retention, so the wedge is a set of retention controls rather than code that never leaves the machine. Native MCP support rides the open-tooling wave. Market timing is late (Claude Code and Codex are entrenched), and price is now the clearer wedge. The weakness is the v0.1 reality: a sharp positioning story needs the product maturity to back it, and right now the integrations and benchmark proof lag the narrative.

“At $1/$2 it's a fraction of Claude Sonnet or Opus per token — the catch is retries on hard tasks can eat the savings.”

The unit economics are strong. At $1.00 / $2.00 per 1M tokens, Grok Build is materially cheaper than Claude Sonnet 4.6 ($3/$15) and roughly 10x cheaper than Opus 4.7 per token — meaningful for coding agents that burn tens of millions of output tokens per project. The CLI bundled into SuperGrok ($30/mo) or X Premium+ ($40/mo) undercuts stacking Cursor + Claude Code subscriptions. The financial caveat: lower benchmark quality means more retries and more developer-supervision time on hard problems, which can quietly erode the per-token savings. For routine refactors and well-scoped tasks, the economics clearly favor Grok Build; for hard, novel problems, factor in the human-supervision overhead.

“Plan mode and 8 worktree sub-agents are a genuine win — the CLI just feels less polished than Claude Code, and 256K bites on big repos.”

For builders, the standout UX is plan mode (approve a step list before any mutation) and parallel sub-agents isolated in Git worktrees, which make multi-file refactors fast and reviewable. Native MCP support plugs into the growing tool ecosystem, and function calling / structured outputs are solid via the OpenAI-compatible interface. The friction: the CLI is less polished than Claude Code — fewer editor integrations and plugins — always-on reasoning means waiting on each action, and the 256K context forces manual context selection on real monorepos that a 1M-context tool avoids. For well-scoped tasks it's productive; for sprawling repos it asks more of the developer.

“As a daily coding driver it's serviceable and cheap, but I still reach for Claude Code on anything hard — and I feel the reasoning pause.”

For a developer living in the CLI day to day, Grok Build is serviceable and notably cheap, with plan mode adding a reassuring approval step before edits. On routine tasks it ships fine. But on hard problems the quality gap to Claude shows, the always-on reasoning pause on every action is noticeable, and the smaller context means more time spent curating what the agent sees. For proprietary code, turning off code-data retention in /privacy is worth doing on day one. As an everyday tool it earns a spot in the rotation; as the single tool for all coding, most heavy users will still keep Claude Code alongside it.

“One self-reported SWE-bench number on xAI's own harness, no other benchmarks, v0.1 — the local-first story doesn't hold, and the quality story is unproven.”

Adversarially, Grok Build 0.1's evidence base is thin even by xAI standards: a single headline benchmark (70.8% SWE-bench Verified) measured on xAI's own internal harness rather than an independent run, and aggregators showing zero sourced evals. The 70.8% also trails Claude Sonnet, so the one number on offer is not a winning one. v0.1 versioning is honest signaling that this is early. The local-first claim, once the strongest part of the pitch, does not survive xAI's own security FAQ: the CLI retains code data unless you turn that off or enable Zero Data Retention. But "credible Claude Code competitor" rests largely on positioning, not on independently verified coding quality, and the November 2024 cutoff is a real handicap for current frameworks.