Grok Build API — Cursor Workflow Data Is xAI's Real Bet

xAI made grok-build-0.1 available through the public API on May 29 — no X Premium+ subscription, no SuperGrok gate, just an OpenAI-compatible endpoint at api.x.ai/v1 with $25 in free credits at signup. That alone would be a minor story: another coding model enters the API market. What makes this worth your attention is what dropped the following week. On June 5, Musk confirmed Git worktrees support for parallel agent branches and, more importantly, that V9-Medium — a 1.5-trillion-parameter foundation model, three times the current production size — was trained on real Cursor developer workflow data under the SpaceXAI–Anysphere deal. The API access democratizes the tool. The Cursor data pipeline is the thing you should actually care about.

TL;DR

• What: grok-build-0.1 is now publicly available via the xAI API at $1/M input, $2/M output tokens — no subscription required
• Why it matters: V9-Medium (1.5T params) was trained on Cursor developer workflow data, not just public GitHub — a fundamentally different training corpus than any competitor
• The deal: SpaceXAI holds a $60B option to acquire Anysphere (Cursor’s parent), structuring permanent ownership of the data pipeline
• Action: Wait for independent V9-Medium benchmarks before making routing decisions — current grok-build-0.1 scores 70.8% on SWE-bench Verified vs. Claude Opus 4.7’s 87.6%

Grok Build Goes Public — What Happened

The model itself is straightforward. grok-build-0.1 is a coding-specific model trained for agentic tasks: web development, debugging, refactoring, with native MCP support. It runs at 100+ tokens per second, carries a 256K-token context window, and is priced at $1 per million input tokens and $2 per million output tokens. That pricing undercuts Claude Sonnet and sits roughly in line with GPT-4.1 mini — cheap enough to be interesting for high-volume agentic pipelines where you’re burning through tokens on speculative branches.

The more significant update came on June 5. Musk posted two confirmations: first, that Grok Build now supports Git worktrees, allowing up to eight parallel sub-agents to work on separate branches from a single repo without touching each other’s code. Second, and far more consequential, that V9-Medium had completed training with Cursor developer workflow data as a core component of its training corpus.

# Calling grok-build-0.1 — OpenAI-compatible endpoint
curl https://api.x.ai/v1/chat/completions \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-build-0.1",
    "messages": [{"role": "user", "content": "Refactor this function to use async/await"}]
  }'

Why This Matters

Every major coding model — Claude Code, Codex, Gemini Code Assist — trains primarily on public GitHub repositories, documentation, and Stack Overflow-style Q&A. That corpus is enormous but it captures finished artifacts: committed code, merged pull requests, polished READMEs. It does not capture the messy, iterative process of how experienced developers actually debug and refactor in real time.

Cursor sessions are different. They contain the full decision chain: the initial prompt, the rejected suggestion, the correction, the follow-up question that narrows the intent, the accepted diff. This is workflow data — the cognitive process of skilled developers working through ambiguous problems, not just the cleaned-up output they push to a branch. Training on this corpus is a bet that the process signal is as valuable as the artifact signal. If that bet pays off, V9-Medium won’t just generate better code completions; it will generate better sequences of actions, which is exactly what matters for agentic use cases where the model needs to plan, execute, evaluate, and retry.

The deal mechanics reinforce how seriously SpaceXAI is taking this bet. The April 2026 agreement gives SpaceXAI the option to acquire Anysphere — Cursor’s parent company — for $60 billion, with an alternative path of paying $10 billion for ongoing collaboration. This is not a one-off data licensing arrangement. It’s structuring to own the corpus permanently. If the acquisition closes, xAI would control both the model and the richest source of developer intent data in the industry, a vertical integration play that no other AI lab can currently replicate.

V9-Medium has not shipped yet. Everything about its capabilities is based on training completion announcements and Musk’s characterization of evaluation results as positive. No independent benchmarks exist. The mid-June release window (extrapolated from “2-3 weeks” stated May 25) may slip. Do not make infrastructure decisions based on projected performance.

The competitive context matters. Current grok-build-0.1 scores 70.8% on SWE-bench Verified. Claude Opus 4.7 sits at 87.6%. That’s a 16.8-point gap — not the kind of margin you close with incremental tuning. The hypothesis is that Cursor workflow data provides a training signal qualitatively different enough to produce a step-function improvement. It’s a reasonable hypothesis. It is not a proven result.

For teams already running multi-model agent pipelines, the pricing angle alone justifies attention. At $1/$2 per million tokens, grok-build-0.1 costs roughly 60–70% less than Claude Sonnet for equivalent context windows. If your pipeline uses a capable primary model for planning and a cheaper secondary model for execution — code generation, test writing, boilerplate — current Grok Build is already a candidate for the secondary slot, benchmark gap notwithstanding. The Git worktrees feature strengthens this case: eight parallel sub-agents working on isolated branches is genuinely useful architecture for decomposed coding tasks, and it ships today, not mid-June.

If you’re evaluating Grok Build as a secondary agent in a multi-model pipeline, test it on your actual codebase now — not on benchmarks. SWE-bench measures resolution of GitHub issues in open-source projects. Your internal refactoring tasks, your domain-specific patterns, your framework conventions may produce very different results.

The Cursor SDK lock-in question cuts both ways here. If SpaceXAI acquires Anysphere, Cursor’s developer data becomes proprietary training fuel for a competing model ecosystem. Engineers who chose Cursor for its editor quality may find their workflow data training a model they don’t use. That’s not necessarily a reason to stop using Cursor — the product remains excellent — but it’s a dependency worth tracking, especially if your organization has data governance requirements around how developer activity data is used.

The Take

I’ve been skeptical of Grok Build since the arena-mode beta — novelty pricing around a model that sits nearly 17 points behind Claude Opus 4.7 on the benchmark that matters most for agentic coding. That skepticism hasn’t changed for the current model. But the V9-Medium training approach is a genuinely different bet, and I think it’s the right one.

Public GitHub data is a commodity at this point. Every lab has it, every lab trains on it, and the performance ceiling from that corpus is approaching. Cursor workflow data — real debug sessions, real refactoring decisions, real intent-to-action chains from experienced developers — is not a commodity. It’s proprietary, it’s high-signal, and xAI is structuring a $60 billion deal to own it permanently.

Here’s my concrete recommendation: do not route primary agentic tasks to grok-build-0.1 today. The 70.8% SWE-bench score makes it unsuitable as a lead agent when Claude Opus 4.7 exists. But do set it up as a secondary execution model in your pipeline — the $1/M pricing makes it nearly disposable for parallel branch work, and the worktrees support is production-ready now. When V9-Medium ships and independent benchmarks land, you’ll have the integration already built and can evaluate a model swap in hours, not weeks. That’s the position you want to be in: ready to move fast on data, not on hype.

Related

• xAI Grok Build Early Beta — Arena Mode — our coverage of the initial beta launch and pricing analysis
• Cursor SDK and Agent Infrastructure Lock-In — the dependency implications of building on Cursor’s ecosystem
• Kimi Code K26 — Claude Code’s Cost Threat — how budget coding models are reshaping multi-model pipeline economics