Grok 4.7: xAI's New Flagship With Longer Reasoning

xAI released Grok 4.7 this week — a roughly 2.1-trillion-parameter model with extra training on SpaceX engineering data, designed for long, multi-step jobs like coding. The price is unchanged from its predecessor. Here is what is new, who it is for, and how it compares in the busiest model-release week of the year.

What actually launched

Grok 4.7 arrived around September 21–22, 2026, hours apart from OpenAI's GPT-6 Sol/Luna and Anthropic's Claude Opus 5.5. Three things define it:

1. Scale. About 2.1 trillion parameters — a roughly 40% jump over its predecessor — making it xAI's largest model to date.
2. Unusual training data. After initial training finished, xAI added SpaceX's engineering corpus: Starlink telemetry, rocket test logs, and internal engineering documents. The bet is that real-world engineering data makes the model better at rigorous technical reasoning.
3. A new training goal. Instead of just answering questions, Grok 4.7 was trained with extended reinforcement learning to handle tasks that take hours and many steps — and to check its own reasoning along the way.

Think of the difference this way: older models are like brilliant students who blurt out the first answer. Grok 4.7 is trained to be the student who shows their work, double-checks it, and keeps going when the problem takes all afternoon.

The numbers, in plain terms

• Price, input (per 1M tokens, under 200K prompt tokens) · Grok 4.6: $2.00 · Grok 4.7: $2.00 (unchanged)
• Price, output (per 1M tokens) · Grok 4.6: $6.00 · Grok 4.7: $6.00 (unchanged)
• Parameters · Grok 4.6: ~1.5 trillion · Grok 4.7: ~2.1 trillion (reported)
• xAI's CursorBench 4.0 score (coding) · Grok 4.6: — · Grok 4.7: 46.3%

Where you can use it: Cursor, the Grok API, Grok Build, Visual Studio (via Copilot's model picker), and other third-party platforms.

Unlike its rivals this week, xAI did not cut prices — it is betting that better long-task performance is worth the same price.

What "thinks longer and checks itself" means in practice

Most AI models today are sprinters: great at one question, one answer. Grok 4.7 is built to be a marathoner:

• Coding sessions that span hundreds of steps — writing, testing, fixing, re-testing.
• Agentic workflows where the model uses tools, reads results, and adjusts its plan.
• Self-verification: pausing to re-examine its own reasoning before finalizing an answer, which should mean fewer confident-but-wrong outputs.

For beginners, the practical takeaway is simple: if you ever give an AI a big, messy job ("help me plan this project," "debug this whole program"), models trained like this are the ones most likely to actually finish it.

How it stacks up — honestly

This was a three-way launch week, so comparisons are inevitable:

• On price, Grok 4.7 ($2/$6) sits between GPT-6 Sol ($2/$10) and Luna ($0.10/$0.50) — but price alone says little about which model finishes your task cheapest.
• On independent benchmarks, the gains look modest so far: one widely cited index scores Grok 4.7 at 46 vs. 44 for Grok 4.6. A real step, not a leap.
• On xAI's own benchmarks, the coding numbers look strong — but as always, those are the company's own tests.

What to be careful about

• The 2.1-trillion figure comes from pre-release reporting (including statements attributed to Musk), not a formal xAI specification sheet. Treat it as reported, not confirmed.
• "Trained on SpaceX data" sounds exciting, but its real-world effect is unproven. It is a plausible idea — engineering data should help engineering reasoning — but no independent test has measured the difference yet.
• Every lab's benchmark charts deserve the same skepticism you would give any sales pitch. Wait for independent evaluations before drawing conclusions.

Explain it to a friend: "Elon Musk's AI company released its biggest model ever — trained partly on SpaceX rocket data, built to work on long coding jobs for hours and double-check its own work. Same price as before. The upgrades look solid but not revolutionary, at least until independent testers weigh in."

Try this: When a model claims to be better at "long tasks," test it the honest way: give it a real multi-step job from your own work, check every step, and compare with what you used before. Your own task is the only benchmark that matters to you.