xAI

Grok 4.7

Provider

xAI

Status

Current

Context

500,000 tok

Price

$2 / $6 /MTok

Knowledge

2026-05

Grok 4.7 is xAI’s flagship model, released on 21 September 2026 on the SpaceXAI API, in Cursor and in Grok Build, with gateway availability on OpenRouter, Vercel and Cloudflare (SpaceXAI). It succeeds Grok 4.6 after six weeks. SpaceXAI describes it as built on “a new, larger base model” with a longer reinforcement learning run weighted toward tasks that take many hours to complete; Elon Musk put the base at about 2.1 trillion parameters on 11 September. The release had been expected around 12 September and was delayed on the 11th; on the 13th Musk announced a larger Grok 4.8, which has no release date.

Artificial Analysis scored Grok 4.7 at 46 on its Intelligence Index v4.3 at xhigh effort, two points above Grok 4.6 (44) and one below GPT-5.6 Sol (47). The price per token is unchanged at $2/$6, but the model uses about 81,000 output tokens per index task, against 36,000 for Grok 4.6, so its cost per completed task is higher: $2.73 at high effort and $3.74 at xhigh, against Grok 4.6’s $1.86. SpaceXAI publishes no SWE-bench figure of any kind.

Quick specs

TRY GROK 4.7 →

What’s new in Grok 4.7

Grok 4.6 was a post-training release on Grok 4.5’s foundation. Grok 4.7 changes the base model: SpaceXAI describes “a new, larger base model” trained with a longer reinforcement learning run on a harder mix of tasks, weighted toward problems that take many hours to complete. SpaceXAI says the model works longer on difficult tasks and checks its own work more carefully.

The context window (500K), modalities (text and image in, text out) and price ($2/$6) are the same as Grok 4.6. Reasoning effort now runs from low to xhigh, with high the default. On the Responses API, encrypted reasoning is always returned so multi-turn conversations keep the model’s reasoning state. A Fast variant at twice the standard rates is available in Cursor and Grok Build only.

Benchmark performance

SpaceXAI’s launch table (21 September 2026)

On SpaceXAI’s table, Grok 4.7 leads GPT-5.6 Sol on five of seven shared rows and trails it on DeepSWE v1.1 and HealthBench Professional. Claude Fable 5.1 leads it on CursorBench, Terminal-Bench 4.0, HealthBench, AA-Briefcase and GDPval; Grok 4.7 leads Fable 5.1 on DeepSWE, EEBench and Harvey. SpaceXAI’s Grok 4.6 figures in this table differ slightly from its August launch table (DeepSWE 65.2% here, 65.9% then). These are vendor numbers.

The independent read (Artificial Analysis)

Source: Artificial Analysis and VentureBeat. Artificial Analysis reports AA-Briefcase as the largest gain (+111 Elo), placing Grok 4.7 behind only Claude Opus 5 and Claude Fable 5.1 on that measure. No SWE-bench Pro or Verified figure exists for Grok 4.7 in any harness.

Safety claims

SpaceXAI describes Grok 4.7 as “the strongest model we’ve tested on refusals and jailbreak resistance”, reports 62.4% on the LatchBio biosafety benchmark and a 3.3% refusal rate on risky dual-use prompts in HackerBench v0.3, and says it built “an entirely new safeguard stack”. These are company-reported; no separate system card was published at launch.

Pricing breakdown

Cost per completed task

Per completed task, Grok 4.7 at xhigh costs more than GPT-6 Astra, and Grok 4.6 remains the cheaper Grok.

How to access Grok 4.7

Via API

Grok 4.7 is generally available as grok-4.7 on SpaceXAI’s API (console.x.ai), with a US regional endpoint at a 10% premium, and through OpenRouter, Vercel and Cloudflare.

from openai import OpenAI
client = OpenAI(base_url="https://api.x.ai/v1", api_key="XAI_API_KEY")

resp = client.chat.completions.create(
    model="grok-4.7",
    messages=[{"role": "user", "content": "Your prompt here"}],
)
print(resp.choices[0].message.content)

Via Cursor and Grok Build

Grok 4.7 launched in Cursor and Grok Build, including Grok Build’s free tier (standard variant only). SpaceXAI did not state which Grok app subscription tiers include it at launch; see the Grok app page and the xAI provider page.

How Grok 4.7 compares

vs Grok 4.6

Grok 4.7 scores two points higher on the Artificial Analysis index (46 to 44), nine points higher on its Coding Agent Index, and leads Grok 4.6 on every row of SpaceXAI’s table, at the same per-token price. It uses about 2.25 times as many output tokens per index task, so a completed task costs $2.73 at high effort against Grok 4.6’s $1.86.

vs GPT-5.6 Sol

GPT-5.6 Sol is one point higher on the index (47 to 46) and publishes a SWE-bench Pro figure (64.6%, vendor) where Grok 4.7 has none. On SpaceXAI’s table Grok 4.7 leads Sol on CursorBench 4.0, Terminal-Bench 4.0, EEBench, AA-Briefcase and Harvey, and trails on DeepSWE and HealthBench. Sol costs less per completed index task ($1.99) despite its higher per-token price.

vs Claude Fable 5.1

Claude Fable 5.1 scores 53 on the index and leads Grok 4.7 on Terminal-Bench 4.0 by about 20 points on SpaceXAI’s own table (57.9% to 38.0%). Fable 5.1 costs $10/$50 per million tokens and $7.63 per index task.

Known limitations

No SWE-bench figure. Neither SpaceXAI nor any standardised harness has published a SWE-bench Pro or Verified result.

Higher token use. About 81,000 output tokens per Artificial Analysis index task, 125% more than Grok 4.6 and 196% more than GPT-6 Astra, which raises cost per task despite the unchanged price.

Long-context calls double the price. Prompts above 200K input tokens are billed at $4/$12.

Fast variant is not on the API. It is limited to Cursor and Grok Build.

Consumer and EU availability not stated. SpaceXAI’s launch material does not list Grok app subscription tiers or EU availability.

Version history

Musk announced Grok 4.8, a model of about 2.5 trillion parameters, on 13 September 2026; it has no release date.

Frequently asked questions

When was Grok 4.7 released?

Grok 4.7 was released on 21 September 2026 on the SpaceXAI API, in Cursor and in Grok Build, with gateway access through OpenRouter, Vercel and Cloudflare. It had been expected around 12 September; Musk delayed it on 11 September.

How much does Grok 4.7 cost?

$2.00 per million input tokens and $6.00 per million output tokens, the same as Grok 4.6. Prompts above 200,000 input tokens are billed at doubled rates, the US regional endpoint costs 10% more, and the Grok 4.7 Fast variant, available only in Cursor and Grok Build, costs twice the standard rates. Artificial Analysis measured $2.73 per completed Intelligence Index task at high effort and $3.74 at xhigh.

How does Grok 4.7 score on benchmarks?

Artificial Analysis scored it 46 on its Intelligence Index v4.3 at xhigh effort, against 44 for Grok 4.6, 47 for GPT-5.6 Sol and 53 for Claude Fable 5.1 and GPT-6 Astra. It measured 56 on the Coding Agent Index with Grok Build and 33% on Terminal-Bench 4.0. SpaceXAI’s own table reports 71.0% on DeepSWE v1.1 and 38.0% on Terminal-Bench 4.0. No SWE-bench figure has been published.

What is Grok 4.7’s context window?

500,000 tokens, with text and image input and text output, and a May 2026 knowledge cutoff.

How is Grok 4.7 different from Grok 4.6?

Grok 4.7 uses a new, larger base model and a longer reinforcement learning run; Grok 4.6 was a post-training release on Grok 4.5’s base. Grok 4.7 scores higher on every published benchmark at the same per-token price, but uses about 2.25 times as many output tokens per task, so its cost per completed task is higher ($2.73 at high against $1.86).

Where does Grok 4.7 rank?

It enters our best AI models ranking at #8, below GPT-5.6 Sol and above Grok 4.6, which keeps the best-value slot on cost per task.

Written at launch, 22 September 2026. Artificial Analysis figures are independent; benchmark table rows are SpaceXAI’s. Pricing and availability reflect the 21 September 2026 launch documentation.