Grok 4.7 Guide: Features, Pricing, Benchmarks & Real-World Performance

Grok 4.7 Comprehensive Guide: Features, Benchmarks, Pricing, and Real-World Testing

On September 21, 2026, SpaceXAI (xAI) officially launched Grok 4.7, its newest flagship model optimized for complex coding, agentic software engineering, and professional knowledge work. Built upon a larger base model architecture than its predecessor, Grok 4.6, Grok 4.7 underwent an extended reinforcement learning training regimen specifically targeted at multi-hour, multi-step problem solving.

The model maintains a 500,000-token context window, a knowledge cutoff of May 2026, and multimodal input capabilities (text and image understanding) paired with text generation. By combining self-verification mechanisms with competitive API pricing ($2.00 per million input tokens and $6.00 per million output tokens for standard requests), Grok 4.7 positions itself as an accessible frontier model for technical teams and enterprise developers.

---

1. Model Architecture, Training, and Core Specifications

Grok 4.7 was trained using an updated reinforcement learning framework weighted toward complex, multi-hour technical tasks. Key structural features include:

• Configurable Reasoning Effort: Developers can specify the reasoning depth using four parameters: low, medium, high (default), and xhigh. Because reasoning cannot be disabled, all responses include reasoning tokens in the total billed output.

• Tiered API Pricing:

  • Standard Requests (<200,000 prompt tokens): Billed at $2.00/1M input tokens, $0.50/1M cached input tokens, and $6.00/1M output tokens.

  • Long-Context Requests (≥200,000 prompt tokens): When a prompt reaches or exceeds 200,000 tokens, all tokens in the request shift to the long-context rate: $4.00/1M input, $1.00/1M cached input, and $12.00/1M output.

  • US Regional Endpoint: Routing requests through the US regional endpoint adds a 10% premium to base token charges.

• Grok 4.7 Fast: Served exclusively within Cursor and Grok Build (and omitted from the public xAI API), Grok 4.7 Fast delivers twice the output generation speed at double the standard token price ($4.00 input / $1.00 cached / $12.00 output per 1M tokens for short contexts).

+-----------------------------------------------------------------------------------+
|                            Grok 4.7 Token Pricing Tiers                           |
+------------------------------------+------------------+----------------+----------+
| Request Category                   | Input / 1M       | Cached / 1M    | Output/1M|
+------------------------------------+------------------+----------------+----------+
| Standard API (<200K Tokens)        | $2.00            | $0.50          | $6.00    |
| Long-Context API (>=200K Tokens)   | $4.00            | $1.00          | $12.00   |
| Fast Variant (Cursor / Build)      | $4.00            | $1.00          | $12.00   |
+------------------------------------+------------------+----------------+----------+

---

2. Benchmark Performance & Independent Evaluations

Benchmarking reveals strong technical performance in domain-specific tasks alongside nuanced trade-offs in raw agentic terminal execution.

+-----------------------------------------------------------------------------------+
|                   Grok 4.7 Benchmark Comparison Across Frontier Models            |
+----------------------------------+----------+----------+--------------+-----------+
| Benchmark Evaluation             | Grok 4.7 | Grok 4.6 | GPT-5.6 Sol  | Fable 5.1 |
+----------------------------------+----------+----------+--------------+-----------+
| CursorBench 4.0 (Coding)         | 46.3%    | 40.4%    | 41.7%        | 51.8%     |
| DeepSWE v1.1 (Software Eng.)     | 71.0%*   | 65.2%    | 72.7%        | 70.0%     |
| EEBench (Electrical Eng.)        | 64.0%    | 53.0%    | 39.4%        | 56.4%     |
| AA Briefcase v1.1 (Elo)          | 1,657    | 1,546    | 1,487        | 1,678     |
| Terminal-Bench 4.0 (xAI-reported)| 37.6%    | 20.3%    | 37.3%        | 57.9%     |
| Harvey Legal Agent Benchmark     | 19.6%    | 15.8%    | 2.5%         | 6.7%      |
| HealthBench Professional         | 56.7%    | 48.5%    | 60.5%        | 62.1%     |
+----------------------------------+----------+----------+--------------+-----------+
* Note: Evaluated at high/xhigh reasoning effort settings.

---

• Artificial Analysis Intelligence Index: Grok 4.7 scored 46 on the Artificial Analysis Intelligence Index, placing SpaceXAI among the top four AI labs globally.

• Agentic Work & Elo Ratings: On private benchmarks for long-horizon professional work, Grok 4.7 achieved 1657 Elo on AA-Briefcase (+111 Elo over Grok 4.6) and 1695 Elo on GDPval-AA.

• Coding Agent Index: Paired with Grok Build, Grok 4.7 scored 56 on the Coding Agent Index (up from 47 with Grok 4.6).

• Token Usage & Speed: Independent measurements recorded an output speed of approximately 188 tokens/second. However, running at xhigh effort consumed ~81,000 output tokens per index task, compared to 36,000 tokens for Grok 4.6 high.

• Hallucination Reductions: Measured hallucination rates on the AA-Omniscience benchmark dropped to 29%, down from 34% in Grok 4.6.

• Terminal-Bench Discrepancies: Independent evaluations by Artificial Analysis and technical analysts recorded Terminal-Bench scores between 26% and 33%, highlighting a performance gap against Claude Fable 5.1 (55%) and GPT-6 Astra (60%) in real terminal environments.

---

3. Real-World Testing & Practical Use Cases

Hands-on technical testing highlights Grok 4.7's performance across live software stacks, vision diagnostics, legal analysis, and multilingual tasks:

A. Full-Stack Agentic Debugging

• Test Bed: A containerized full-stack application (Dirt Dynasty) running Postgres, FastAPI, Nginx, Redis, and Docker Compose.

• Scenario: The application contained an unscripted logic bug in its aggregation logic where the leaderboard silently dropped the fourth judge's score across all entries, altering the winner.

• Result: Operating autonomously via the Hermes agent harness without hints, Grok 4.7 analyzed the multi-container data flow, isolated the scoring aggregation error, patched the backend code, verified the fix, and updated the leaderboard correctly.

B. Electrical Engineering & Visual Circuit Diagnosis

• Scenario: Evaluated on a hand-drawn schematic displaying a 24V power supply, parallel and series resistors, an LED, and a multimeter reading 0V with a misleading label marked "open circuit".

• Result: Grok 4.7 traced the circuit connections pixel-by-pixel, calculated voltage drops, ignored the misleading label, and correctly identified that the primary fault was a reverse-biased LED.

C. Legal Analysis & Conflict Navigation

• Scenario: Evaluated on a complex legal problem involving multi-jurisdictional marriages, alleged bigamy, workplace disputes, and financial claims.

• Result: The model structured its response around legal validity, criminal exposure, and property rights. It surfaced advanced legal concepts - such as putative spouse rights, kinship-based marriage voidance, and conflict-of-laws principles - while avoiding demographic shortcuts or bias.

D. Multilingual Scripting & Logic Boundaries

• Scenario: Requested to build an animated latte-brewing machine in HTML/JavaScript and render the word "coffee" in native scripts across 80 languages, including one fake test language.

• Result: Rendered a functional animation on the first pass. Across 80 languages, it correctly traced the "qahwa" linguistic root across native scripts. When encountering the fake language, Grok 4.7 recognized the absence of a real-world translation and generated a plausible invented term rather than hallucinating a false factual lookup.

---

Also Read this: Grok 4.5: The Complete Guide to SpaceXAI’s Opus-Class Coding Agent Alternatives

---

Step-by-Step Guide: Integration & Access Pathways

Pathway 1: Access via Cursor

1. Open Cursor and navigate to the model picker.

2. Select grok-4.7 or grok-4.7-fast.

3. Verify your subscription's usage limits and token allowances.

Pathway 2: Access via Grok Build CLI

1. Install the CLI using the official terminal command:

curl -fsSL https://x.ai/cli/install.sh | bash

1. Launch Grok Build; Grok 4.7 operates as the default underlying engine.

Pathway 3: Access via xAI API

1. Generate an API key inside the xAI Console (console.x.ai).

2. Set the model parameter to grok-4.7.

3. Send requests to https://api.x.ai/v1/responses or the Chat Completions endpoint.

Example Python API Request:

import os
import requests

api_key = os.getenv("XAI_API_KEY")
url = "https://api.x.ai/v1/responses"

headers = {
    "Authorization": f"Bearer {api_key}",
    "Content-Type": "application/json"
}

payload = {
    "model": "grok-4.7",
    "reasoning": {"effort": "low"},
    "input": "Explain why sorting numerical arrays in JavaScript requires a numeric comparator function."
}

response = requests.post(url, headers=headers, json=payload)
print(response.json())

Note: Setting reasoning effort to low reduces latency for straightforward queries*.*

---

Best Practices

• Match Reasoning Effort to Task Complexity: Use low or medium effort for simple code generation, documentation, and data extraction. Reserve xhigh for complex multi-file refactoring, circuit analysis, or legal synthesis.

• Maintain Conversation Cache Keys: Set a consistent prompt_cache_key (Responses API) or x-grok-conv-id (Chat Completions) across conversation turns to leverage cached prompt discounts ($0.50 / $1.00 per 1M tokens).

• Preserve Encrypted Reasoning State: When managing multi-turn state on the Responses API, pass the returned reasoning.encrypted_content object back unchanged in the next request's input array to maintain reasoning continuity.

---

Common Mistakes

• Expecting Direct Image Generation Output: Grok 4.7 processes image inputs but outputs text only. Image or video creation requires dedicated tools like Aurora or Grok Imagine.

• Unintended Long-Context Escalation: Sending prompts at or above 200,000 tokens automatically doubles input, cached, and output token rates for the entire request. Compact long histories when feasible.

• Attempting Batch API Calls: Grok 4.7 does not currently support the xAI Batch API.

• Targeting Fast via the Public API: Grok 4.7 Fast is not exposed on the public xAI API; it is restricted to Cursor and Grok Build environments.

---

Practical Examples

Example: Automated Bug Diagnostics with Cost Tracking

When submitting codebases for automated debugging, inspect the response's usage object to monitor exact expenditure via the cost_in_usd_ticks field.

{
  "model": "grok-4.7",
  "usage": {
    "prompt_tokens": 12450,
    "completion_tokens": 820,
    "cost_in_usd_ticks": 29820
  }
}

Tracking cost_in_usd_ticks allows teams to log exact per-request costs across high-volume pipelines*.*

---

Frequently Asked Questions (FAQ)

Q1: When was Grok 4.7 released?

Grok 4.7 was officially released on September 21, 2026, with developer documentation and API availability going live the same day.

Q2: How much does the Grok 4.7 API cost?

For prompts under 200,000 tokens, standard rates are $2.00 / 1M input tokens, $0.50 / 1M cached input tokens, and $6.00 / 1M output tokens. For prompts at or above 200,000 tokens, rates double to $4.00 / $1.00 / $12.00 per 1M tokens.

Q3: What is Grok 4.7's context window?

Grok 4.7 features a 500,000-token context window.

Q4: Can Grok 4.7 generate images or videos?

No. Grok 4.7 accepts text and image inputs and produces text outputs. Media generation is handled by separate xAI tools such as Aurora or Grok Imagine.

Q5: What is Grok 4.7 Fast?

Grok 4.7 Fast is a variant served on higher-speed infrastructure that delivers twice the output generation speed at double the standard token cost. It is available only in Cursor and Grok Build.

Q6: Can reasoning be turned off in Grok 4.7?

No. Reasoning effort can be configured (low, medium, high, xhigh), but it cannot be disabled entirely. Reasoning tokens are billed as part of output usage.

---

Key Takeaways

• Targeted Domain Strengths: Grok 4.7 leads among compared models on the Harvey Legal Agent Benchmark and electrical engineering visual diagnostics (EEBench).

• Competitive Pricing: At $2.00/$6.00 per million tokens, Grok 4.7 offers frontier capabilities at a lower cost than comparable models like Claude Opus 5.1 or GPT-5.6 Sol.

• Substantial Context & Modality: Features a 500,000-token context window with text and image understanding.

• High Token Output at High Reasoning: Extended self-verification loops increase total output token volume during xhigh reasoning tasks.

---

Conclusion

Grok 4.7 represents a cost-effective option for developers, legal analysts, and engineering teams seeking strong technical reasoning and self-verification without paying premium API prices. While independent benchmarks highlight performance gaps in raw terminal execution compared to higher-cost rivals, its pricing structure and multimodal capabilities make Grok 4.7 a practical choice for high-volume agentic software development and technical workflows.

---

Authoritative External References

1. SpaceXAI Official Announcement: x.ai/news/grok-4-7

2. SpaceXAI Developer Documentation: docs.x.ai

3. Artificial Analysis Grok 4.7 Benchmark Index: artificialanalysis.ai

4. MindStudio Technical Review: mindstudio.ai

5. EvoLink API Documentation: evolink.ai

---

Also Read this: Grok 4.5: The Complete Guide to SpaceXAI’s Opus-Class Coding Agent Alternatives

---