How do you use Grok 4.5?

Grok 4.5 is the latest large language model from SpaceXAI, the company behind the Grok chatbot, whose earlier models came out under the xAI name. It arrived on July 8, 2026, and the company describes it as its strongest model so far, aimed at three kinds of work: writing and fixing code, running agentic tasks (multi-step jobs where the model uses tools on its own), and everyday knowledge work such as spreadsheets and documents. It was trained together with Cursor, the AI code editor.

Two claims drive the interest. Elon Musk called it "an Opus-class model, but faster, more token-efficient and lower cost," comparing it with Anthropic's top Claude models. And the price is low: $2 for every million tokens sent in and $6 for every million generated. A token is a small piece of text, roughly three-quarters of an English word, and it is the unit AI providers bill by. For comparison, TechCrunch lists Claude Opus 4.7 at $5 and $25 for the same amounts.

Where you can use it today

SpaceXAI's announcement names three entry points:

SpaceXAI is also offering free use of Grok 4.5 in Grok Build and Cursor for a limited period, which makes this a good week to experiment without a bill.

Calling it from your own code

If you build with AI, the API route takes a few lines. You create an API key in the SpaceXAI console, store it as an environment variable, and send a request to the responses endpoint:

import os, requests

reply = requests.post(
    "https://api.x.ai/v1/responses",
    headers={"Authorization": f"Bearer {os.environ['XAI_API_KEY']}"},
    json={
        "model": "grok-4.5",
        "input": "This Python function should return the average of a list "
                 "but crashes on an empty list. Fix it and explain the fix "
                 "in two sentences:\n\ndef average(xs):\n    return sum(xs) / len(xs)",
    },
    timeout=120,
)
print(reply.json())

Print the whole response the first time. It shows the answer and, in its usage section, the token counts you will need for the comparison below.

A worked side-by-side test

Benchmarks published at launch are the company's own. On its charts, Grok 4.5 is close to the leaders on several coding tests but mostly just behind the best score, and TechCrunch describes it as competitive yet short of best in class. Your own task matters more than any chart. Here is a test you can run in about an hour.

1. Pick three real tasks you have already solved, so you know the right answer: for example, fixing a known bug, writing a small function with tests, and turning a messy CSV into a clean summary table.
2. Write one prompt per task and send exactly the same prompt to Grok 4.5 and to the model you use now.
3. Score each answer on four points: correct or not, how many fixes you had to make, tokens used (from the API response), and time to finish.
4. Repeat each task twice, because the same model can give different answers on different runs.

Your results sheet might look like this. The figures are illustrative, not measured:

A table like this tells you something no launch post can: whether lower cost and speed hold up on your work, and where the cheaper model needs extra checking.

How does it compare for current events?

Start with what the model is for. Grok 4.5's announcement is about coding, agents and office documents. It says nothing about news or real-time information, so there is no published evidence yet on how it handles current events.

For any model, knowledge of recent events comes from search tools, not from the model itself. A model only knows what was in its training data, which stops at a fixed date. Grok's advantage has long been its tools: when SpaceXAI, then xAI, launched Grok 4, it trained the model to choose its own search queries for the web and for X, and its API offered live search across X, the web and news. Whether you get fresh information therefore depends on whether search is switched on in the product or API call you use, not only on the model name.

To compare fairly, ask the same question about something from this week in Grok and in another assistant with search turned on, and require dates and links in the answer. Then open the links. Check that each claim appears in the source it cites, and that the date matches. Posts on X arrive fast but are often unverified, so treat a summary built mostly from posts as a lead to check, not a finished answer.