Grok 4.7: the technical guide to benchmarks, API, pricing and how it compares

01What is Grok 4.7

Grok 4.7 does not build on Grok 4.6's base; it runs on a new, larger base model. The developer, now operating as SpaceXAI (formerly xAI), lists four main changes. (MarkTechPost)

1. A new, larger base model instead of the base carried over from Grok 4.6.
2. Longer reinforcement learning (RL) focused on tasks that take many hours.
3. Better self-checking and long-context work: the model verifies its own work more carefully.
4. Native support for the Grok Bot environment for conversational and knowledge work.

According to Elon Musk, Grok 4.7 has 2.1 trillion parameters, 40% more than Grok 4.6 at 1.5 trillion. That comes from a post on X, not from official technical documentation. (Basenor)

On September 23, Musk posted on X that Grok 4.7 has been climbing AI model leaderboards since launch; Basenor attributes this to post-launch optimizations. Musk did not share specific rankings or benchmarks. (Basenor)

02The Grok family and xAI products

Grok 4.7 replaces Grok 4.6 as the main model and is the default model in the Grok Build tool.

Grok Bot for Enterprise is free for Grok and Cursor business customers. (The Soo Group)

03Grok 4.7 technical specifications

Grok 4.7 has a 500,000 token context window, text and image input and four reasoning levels.

Grok 4.7 vs Grok 4.6

Grok 4.7 improved on every row, most of all on terminal tasks.

04How to use Grok 4.7

Grok 4.7 is available in the Grok API, in Cursor on all plans, in Grok Build (as the default model) and through OpenRouter, Vercel and Cloudflare. (MarkTechPost)

Through the xAI API

A Python example with the official xai_sdk, following MarkTechPost:

import os
from xai_sdk import Client
from xai_sdk.chat import user

client = Client(api_key=os.getenv("XAI_API_KEY"))
chat = client.chat.create(model="grok-4.7")
chat.append(user("Explain the structure of this repository."))
print(chat.sample().content)

The model also supports the Responses API and Chat Completions, so it works with OpenAI-compatible libraries. Check the exact parameters (reasoning level, tools) in the xAI documentation.

Practical API tips

• Caching: xAI recommends setting prompt_cache_key in the Responses API for reliable cache hits (in Chat Completions, use the x-grok-conv-id header); cached input costs $0.50 instead of $2.
• Data in the U.S.: the regional endpoint https://us.api.x.ai/v1 keeps inference in the U.S. for a 10% surcharge.
• Speed: Grok 4.7 Fast (2x faster output, 2x the price for prompts under 200K tokens) is only in Cursor and Grok Build, not in the public API.

Sources: MarkTechPost, xAI: pricing

05Grok 4.7 benchmarks and how it compares

In xAI's table, Grok 4.7 leads on domain agent tests (EEBench, legal) and edges out Claude Fable 5.1 on DeepSWE v1.1. On CursorBench and terminal tasks, though, it stays behind Fable 5.1. The table from xAI's announcement:

Fable 5.1 leads on 4 of 7 benchmarks, but it costs 5x more on input and about 8.3x more on output. On the cost per task chart for CursorBench 4.0, xAI places Grok 4.7 on the best price-to-performance frontier. (MarkTechPost) For more on the competitors, see our Claude Fable 5.1 guide, GPT-6 Astra guide, Gemini 3.8 Flash guide and Muse Spark 1.3 deep dive.

xAI left GPT-6 Astra out of its main table. In a side chart of the same announcement, though, Astra leads EEBench with 69.3% ahead of Grok 4.7 (64%), while trailing on GDPval (1542 vs 1695). (xAI)

Independent evaluation

On the Artificial Analysis Intelligence Index v4.3, Grok 4.7 scores 46, tied with Xiaomi's MiMo-V2.6-Pro and just 2 points above Grok 4.6 (44). For comparison: Claude Opus 5.5 scores 58, and Claude Fable 5.1 and GPT-6 Astra score 53. (Artificial Analysis, as of September 26, 2026) TeamDay recommends Grok 4.7 as the best value for coding. (TeamDay)

For all seven September models side by side, with pricing and benchmark charts, see our big comparison of September 2026 AI models.

06Grok 4.7 safety

Grok 4.7 ships with an entirely new set of safeguards. xAI calls it the most robust model it has tested at refusing harmful requests and resisting jailbreaks. (MarkTechPost)

The safeguards aim to combine strong jailbreak resistance with a low refusal rate for legitimate cybersecurity and biology work. (OpenRouter)

Note: HackerBench is an internal xAI benchmark, and independent safety verification is still missing. We did not find a system card comparable to those from OpenAI or Anthropic in the available sources.

07Grok 4.7 pricing and availability

Grok 4.7 costs the same as Grok 4.6: $2 per million input tokens and $6 per million output tokens. From 200,000 tokens in the prompt, the price doubles.

OpenRouter currently lists a lower price of $1.60 / $4.80 (cache $0.40) and a priority variant at $3.20 / $9.60. (OpenRouter) Check the price with your specific provider before you deploy.

Grok 4.7 Fast costs twice the standard price for prompts under 200,000 tokens ($4 / $12, cache $1) and $6 / $18 from 200,000 tokens (cache $1.50). It is only in Cursor and Grok Build. In Grok Build, Grok 4.7 is also available for free, but the Fast variant is not. (xAI: pricing, Basenor)

Grok 4.7 does not support the Batch API with cheaper bulk processing. (OmniaKey)

08Who should use Grok 4.7

Grok 4.7 pays off for teams that want a strong model for code and agents at a fraction of the price of Anthropic's and OpenAI's flagships.

The recommendations are based on the benchmarks and pricing above. Independent tests so far show improvement, not a clear win over the competition. (GlobalGPT)

09How to prompt Grok 4.7

Grok 4.7 is trained on tasks that last hours and on careful self-checking, so give it a goal and a way to verify the result.

1. Set the reasoning level on purpose. The default is high; save xhigh (which xAI used for its benchmarks) for the hardest tasks.
2. Describe what done looks like. The model checks its own results; the more precise the criteria, the better the self-check.
3. Watch the 200,000 token line. Above it the price doubles, so consider whether you need the whole context or just the relevant parts.
4. Use a stable prefix and `prompt_cache_key`. Put the system prompt and documentation at the start, unchanged, so they come from the cache at a quarter of the price.
5. Turn on tools when you need current data. Web search and X search are built in.
6. Use a JSON schema for structured outputs in response_format; the model supports it. (OpenRouter)

10Grok 4.7 FAQ

11Verdict: is Grok 4.7 worth it?

Grok 4.7 is a solid free upgrade: a bigger, smarter model at the same price as Grok 4.6. For coding and domain agents, it offers one of the best price-to-performance ratios of September 2026.

It is not enough for the top spot, though. Claude Fable 5.1 leads on CursorBench and terminal work, GPT-6 Astra leads on EEBench, and most of the numbers so far come from xAI itself. If you pay per token and want a strong model for everyday development, Grok 4.7 is worth testing; keep a pricier model in reserve for critical tasks.

Want to put AI agents to work in your own processes? We help you choose the model and connect it to your systems as part of our AI implementation service.

12Sources

• xAI: Grok 4.7 announcement
• xAI: Grok 4.7 documentation
• xAI: API pricing
• MarkTechPost: SpaceXAI releases Grok 4.7
• Digital Applied: what changed
• OmniaKey: Grok 4.7 review
• iWeaver: specs and comparison with 4.6
• OpenRouter: Grok 4.7
• Basenor: Grok 4.7 in 5 points
• GlobalGPT: Grok 4.7 review
• TeamDay: best AI models of 2026
• Local AI Zone: September 2026 models (Grok 4.6)
• The Soo Group: September 2026
• Anthropic: Claude Fable 5.1