Updated on September 24, 2026
Grok 4.7 is SpaceXAI's frontier model for coding, agentic tasks and knowledge work, the direct successor to Grok 4.6. It landed one day before Claude Opus 5.5 and GPT-6 Sol, with API rates unchanged at $2 per million input tokens. The context window jumps to 500,000 tokens, with text and image input. You can also try it at no cost inside Grok Build, SpaceXAI's coding agent.
Grok 4.7 runs on a larger base model than Grok 4.6, trained with a longer reinforcement learning run on tasks that take hours to finish. According to SpaceXAI, it checks its own work with greater care, holds very long context and natively understands the Grok Bot harness for conversational work. A rebuilt safeguard stack and cybersecurity capabilities round out the update.
In practice, you call the grok-4.7 identifier, pick one of four reasoning effort levels, and the Responses API returns encrypted reasoning so multi-turn sessions keep their train of thought without extra setup. Function calling, web search, X search and code execution come standard, with a knowledge cutoff of May 2026. Paste a whole repository into the context, describe the bug, and the agent gets going. The model is also available in Cursor, where a Fast variant doubles output speed.
SpaceXAI reports 71.0% on DeepSWE v1.1 and 46.3% on CursorBench 4.0, measured at the xhigh effort setting and published by the company founded in 2023 by Elon Musk under the xAI name. Independent measurements are still the missing piece.
Timing tells its own story. The model shipped one day ahead of Claude Opus 5.5, billed at $4 and $20 per million tokens, and GPT-6 Sol, at $2 and $10. The three do not sit the same exams, so crowning a winner would be premature, and the LLM ranking page helps track the pecking order week after week.
| Benchmark | Reported score | What it measures |
|---|---|---|
| DeepSWE v1.1 | 71.0% | Software engineering |
| CursorBench 4.0 | 46.3% | Long-running coding tasks |
| Terminal-Bench 4.0 | 37.6% | Multi-hour terminal work |
| HealthBench Professional | 56.7% | Professional health questions |
| Harvey Legal Agent Benchmark | 19.6% | Agentic legal work |
The public API bills Grok 4.7 at $2 per million input tokens and $6 for output, identical to Grok 4.6. Grok 4.7 Fast, twice the speed at twice the rate, stays exclusive to Cursor and Grok Build, while Grok Build carries a free tier for the standard model. The grok-4.7 identifier is live on OpenRouter, Vercel and Cloudflare as well. Availability on grok.com and inside X, Premium and SuperGrok included, is not spelled out in the official pages, so check grok.com directly. Rates move around, and SpaceXAI's pricing page is the only figure worth quoting.
For an outside read on real-world speed and cost per finished job, measurement sites such as Artificial Analysis run their own meters (handy, since a cheap token that burns thousands of extra tokens stops being cheap rather quickly).
Grok 4.7 is SpaceXAI's frontier model for coding, agentic tasks and knowledge work, succeeding Grok 4.6. It takes text and images as input, handles a 500,000-token context window, and is served through the Grok API, Cursor, Grok Build, OpenRouter, Vercel and Cloudflare.
On the API, it runs $2 per million input tokens and $6 per million output tokens, the same rates as Grok 4.6. The Fast variant costs double and only lives in Cursor and Grok Build. Consumer subscription prices on grok.com and X are best checked on grok.com directly.
A larger base model, a longer reinforcement learning run on harder tasks, better self-verification and stronger long-context handling, at the same price. SpaceXAI's documentation provides migration guides from Grok 4.6.
Budget and your own test suite should decide. Opus 5.5, released the next day, is billed at $4 input and $20 output per million tokens, against $2 and $6 for Grok 4.7. The two companies publish different benchmarks, so the safest comparison is running both on your own repositories.
Verdict: Two dollars per million input tokens for a frontier model, the math practically writes itself. Cursor regulars, teams running coding agents around the clock and anyone following the model race will want Grok 4.7 on their test bench while independent numbers roll in.
GPT-6 Astra 
Claude Fable 5.1 
Claude Fable 5 
