Grok 4.7 API setup on OpenRouter

One-sentence takeaway

The OpenRouter model page labels x-ai/grok-4.7 as SpaceXAI's Grok 4.7, lists input / output prices of $1.60 / $4.80 per million tokens, and gives OpenRouter SDK and cURL examples; the reasoning-level field in the request body, and the list prices for Low, Medium, and High, were not read in this collection.

Use cases

• Suitable tasks: The page describes the model as a flagship for coding, agentic tasks, and knowledge work, and says it is stronger than the previous Grok 4.6 at long-running software engineering, verifying its own work, long context, and drafting documents and presentations.

• Unsuitable tasks: Not stated

• Applicable model versions: The page title is SpaceXAI: Grok 4.7, and the API model ID is x-ai/grok-4.7. This note records only that ID as it appears on this page.

• Applicable clients, agents, or APIs: The Quick Start says the OpenRouter API is OpenAI-compatible, and shows the OpenRouter SDK and POST https://openrouter.ai/api/v1/chat/completions. The same page also lists POST https://openrouter.ai/api/v1/responses and POST https://openrouter.ai/api/v1/messages. Public traffic sources listed in the Apps section at collection time include pi, Hermes Agent, Claude Code, Codex, and omp.

• Recommended reasoning levels and parameters: The Quick Start examples do not set a reasoning level. In the parameter table, the default for reasoning is —; the stated defaults are temperature 0.7 and top_p 0.95. The page says to turn reasoning on with the reasoning parameter and to read reasoning_details in the response; when continuing a conversation, keep the full reasoning_details. The specific field names inside reasoning are not stated. The level names visible in the Providers filter are All, Low, Medium, High, and Extra high. After Extra high is selected, the query parameter in the address bar is reasoningEffort=xhigh. The query-parameter values for Low, Medium, and High are not stated. The page says the benchmarks SpaceXAI reports use xhigh reasoning effort; that is a benchmark note, not a parameter in the request examples.

Ready-to-use content

Below is the setup already visible in this page's Quick Start. It is not a complete prompt. The API key uses only the placeholder given on the page.

Set the environment variable:

export OPENROUTER_API_KEY=sk-or-v1-...

OpenRouter TypeScript SDK example. The page says to read the stream so that reasoning tokens appear in the usage of the final chunk:

import { OpenRouter } from "@openrouter/sdk";

const openrouter = new OpenRouter({
  apiKey: "<OPENROUTER_API_KEY>"
});

// Stream the response to get reasoning tokens in usage
const stream = await openrouter.chat.send({
  chatRequest: {
    model: "x-ai/grok-4.7",
    messages: [
      {
        role: "user",
        content: "How many r's are in the word 'strawberry'?"
      }
    ],
    stream: true
  }
});

let response = "";
for await (const chunk of stream) {
  const content = chunk.choices[0]?.delta?.content;
  if (content) {
    response += content;
    process.stdout.write(content);
  }

  // Usage information comes in the final chunk
  if (chunk.usage) {
    console.log("\nReasoning tokens:", chunk.usage.completionTokensDetails?.reasoningTokens);
  }
}

cURL from Enable streaming on the same page:

curl -N https://openrouter.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -d '{
  "model": "x-ai/grok-4.7",
  "stream": true,
  "messages": [
    {"role": "user", "content": "Hello"}
  ]
}'

The Get Code block on the page also includes an OpenRouter SDK call with an image input. Python, TypeScript (fetch), Python (OpenAI), TypeScript (OpenAI), and the other language samples in Get Code besides the block below were not opened this time; their contents are not stated.

import { OpenRouter } from "@openrouter/sdk";

const openrouter = new OpenRouter({
  apiKey: "<OPENROUTER_API_KEY>"
});

const response = await openrouter.chat.send({
  model: "x-ai/grok-4.7",
  messages: [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "What is in this image?"
        },
        {
          "type": "image_url",
          "image_url": {
            "url": "https://live.staticflickr.com/3851/14825276609_098cac593d_b.jpg"
          }
        }
      ]
    }
  ]
});

console.log(response.choices[0].message.content);

The Chat Completions endpoint states:

• POST https://openrouter.ai/api/v1/chat/completions

• Authorization: Bearer $OPENROUTER_API_KEY

• Content-Type: application/json

• HTTP-Referer: optional — your site URL, for rankings

• X-Title: optional — your site name, for rankings

• Model: x-ai/grok-4.7

Names, types, and defaults read from the parameter table:

Test or workflow steps

1. Create an API key in the OpenRouter dashboard and set it as the environment variable OPENROUTER_API_KEY, following the page example.

2. Request the model x-ai/grok-4.7. The Chat Completions URL is https://openrouter.ai/api/v1/chat/completions.

3. To watch the response as it arrives, add "stream": true as in the page example. The streaming SDK example reads reasoning tokens from chunk.usage.completionTokensDetails.reasoningTokens on the last chunk.

4. The page requires using reasoning to turn reasoning on, and reading the thinking process from reasoning_details in the response; when continuing across turns, send the full reasoning_details back with the messages. The exact JSON for the request is not stated.

5. The FAQ states that the model accepts tools and tool_choice, and does structured output through a JSON schema in response_format. A schema example is not stated.

Raw evidence and data

The page title is "Grok 4.7 - API Pricing & Benchmarks | OpenRouter", the main heading is "SpaceXAI: Grok 4.7", and the model ID is x-ai/grok-4.7.

Model description after expanding, in the original wording:

Grok 4.7 is SpaceXAI's flagship model for coding, agentic tasks, and knowledge work, succeeding Grok 4.6. It is particul… This is a necessary excerpt; read the original source for full context.

Header price and specs: input / output $1.60 / $4.80 per 1M, context 500K, released Sep 21, 2026.

Prices and capabilities in the FAQ, in the original wording:

• Grok 4.7 costs $1.60/M input tokens and $4.80/M output tokens, with separate rates for Cache Read at $0.40/M tokens and Web Search at $5.00/1K calls.

• Grok 4.7 has a 500,000 token context window.

• Yes. Grok 4.7 accepts tools and tool_choice for function calling. It also supports structured outputs via a JSON schema in response_format.

• Grok 4.7 accepts text, images, and files such as PDFs as input and returns text.

• Grok 4.7 was released on September 21, 2026.

List prices read from the Providers table when the default Reasoning effort is All. Latency, throughput, and uptime are the live figures at that time:

The two Priority rows sit under "Not used in Standard routing". The routing description lists Balanced (price + speed), Nitro (fastest), Floor (cheapest), and Exacto (highest tool-calling accuracy). The list price for each of these modes is not stated.

After Extra high was selected, the URL was https://openrouter.ai/x-ai/grok-4.7?reasoningEffort=xhigh. The prices read that time were still $1.60 / $4.80 / $0.40 for the standard endpoint and $3.20 / $9.60 / $0.80 for the Priority endpoint. That screen did not show the Private / Logs pairing again, so the privacy labels are not stated. The latency snapshot at that time was:

Input, Output, Cache read, latency, throughput, and uptime for the Low, Medium, and High levels are all not stated.

The last complete reading of live effective prices in the Pricing section:

• Weighted Avg Input Price: $0.7196 /M tokens

• Weighted Avg Output Price: $6.002 /M tokens

• SpaceXAI: Effective in $0.687, Effective out $6.27, Cache hit rate 83.3%, Token share 1d 45.2%

• SpaceXAI: Effective in $0.747, Effective out $5.90, Cache hit rate 81.8%, Token share 1d 54.8%

• SpaceXAI Priority: Effective in $0.610, Effective out $0.00, Cache hit rate 83.3%, Token share 1d 0.0%

• SpaceXAI Priority: Effective in $0.734, Effective out $4.45, Cache hit rate 81.5%, Token share 1d 0.1%

These four effective-price rows were not paired again with Private / Logs, so the correspondence is not stated. The page states that caching and discounts make the amount actually paid lower than the list price. This set of effective prices changed during collection; the list above keeps only the last complete reading.

The first full-page reading of the P50 / past-3-days snapshot in the Performance section:

• Throughput: 65 tok/s, P50, best across providers

• Latency: 1.14s, P50, best provider

• Throughput average: SpaceXAI 61 tok/s, SpaceXAI (ZDR) 59 tok/s, SpaceXAI (Priority) 56 tok/s

• Latency average: SpaceXAI 1.21 s, SpaceXAI (Priority) 1.36 s, SpaceXAI (ZDR) 1.39 s

• E2E Latency average: SpaceXAI (Priority, ZDR) 6.12 s, SpaceXAI (ZDR) 7.52 s, SpaceXAI (Priority) 8.34 s

• Tool Call Error Rate average: SpaceXAI (Priority, ZDR) 0.00 %, SpaceXAI (Priority) 0.00 %, SpaceXAI (ZDR) 0.14 %

• Structured Output Error Rate average: SpaceXAI (Priority, ZDR) 0.00 %, SpaceXAI (ZDR) 2.44 %, SpaceXAI 5.13 %

• Cache Hit Rate average: SpaceXAI 81.85 %, SpaceXAI (ZDR) 81.56 %, SpaceXAI (Priority, ZDR) 80.28 %

• Uptime (3d): 100.00%

• Availability (3d): 97.06%, window Sep 19, 11 AM - Sep 22, 11 AM

• OpenRouter Availability on the past-24-hours chart: 97.17%

Whether ZDR and "Privacy: Private" are the same endpoint is not stated; this time the page did not write them as the same label, so the correspondence is not stated.

Token volumes from the first reading of Activity: Prompt 5.16B, Reasoning 81.9M, Completion 34.9M. Public traffic in the Apps section at that time: pi 957M tokens, Hermes Agent 940M tokens, Claude Code 614M tokens, Codex 594M tokens, omp 552M tokens.

The Benchmarks table states the source as Artificial Analysis, and every name includes Grok 4.7 (xhigh):

The page also writes Intelligence Index "Better than 92% of models compared". These scores are Artificial Analysis metrics republished on the page, and the page states that the benchmarks SpaceXAI reports use xhigh.

Limits of applicability

• List prices, cache prices, and the Web Search price come from this one OpenRouter model page. The tiered prices for Low, Medium, and High were not read, and they cannot be extrapolated from the All or Extra high list prices.

• Extra high is confirmed only for the query parameter reasoningEffort=xhigh and the list prices on that screen. How to write Low, Medium, High, Extra high, or xhigh in the request body is not stated.

• temperature 0.7 and top_p 0.95 are the defaults in the parameter table. The Quick Start examples do not pass those two fields, and they do not pass reasoning.

• Effective prices, latency, throughput, uptime, token volumes, and app traffic are a page snapshot from 2026-09-22 and will change on refresh.

• Benchmark scores belong to Artificial Analysis, as the page writes them, and they correspond to xhigh. They cannot be written as OpenRouter's own evaluation across all tasks, and they cannot be written as scores for other reasoning levels.

• The FAQ states that input is text, images, and files such as PDFs, and that output is text. Other modalities are not stated.

• This note did not open the xAI documentation, and it does not count parameters from other Grok models as part of x-ai/grok-4.7.

Reproduction notes

1. Open https://openrouter.ai/x-ai/grok-4.7 . The collection date for this pass is 2026-09-22.

2. The header shows the model ID, $1.60 / $4.80, 500K, and Sep 21, 2026. After expanding Show more, the full model description is readable. After expanding the FAQ, the cache price $0.40/M and Web Search $5.00/1K are readable.

3. On a wide screen, the #api region that holds Quick Start uses a style that hides it on large screens. After the viewport is narrowed to about 820px wide, that region is visible, and the TypeScript SDK, cURL, and parameter table can be read.

4. The Providers default filters are Standard, Reasoning effort All, and Latency / throughput P50. The list prices and privacy labels on the All screen were read. Extra high was read, and the URL contains reasoningEffort=xhigh. Do not fill in numbers from the unopened Low, Medium, and High filters; the list prices for those three levels stay unstated.