What Is Grok? Models, Features, API and Grok 4.5 Explained

Grok is SpaceXAI's family of AI models and the assistant built around them. It can answer questions, reason over complex tasks, search the web and X, work with images and files, and power developer tools through the SpaceXAI API. As of July 23, 2026, grok-4.5 is the flagship API model for coding, agentic tasks, and knowledge work.

This guide separates documented capabilities from marketing claims and rumors. For a deeper technical view, see our Grok 4.5 coding analysis. For costs, go directly to the Grok 4.5 API pricing comparison.

Grok at a glance

How Grok evolved

Grok 3: reasoning and DeepSearch

Grok 3 launched in beta on February 19, 2025. SpaceXAI said it was trained on the Colossus supercluster with ten times the compute used for its previous state-of-the-art models. Grok 3 introduced dedicated reasoning modes and DeepSearch, an agent that searches and synthesizes information from multiple sources.

The official announcement also documented a one-million-token context window for Grok 3. That number belongs to Grok 3; context limits differ by model and should not be carried forward to every later Grok release.

Grok 4: native tool use

Grok 4 arrived on July 9, 2025 with native tool use and real-time search integration. It could decide when to use a code interpreter, web search, or X search while reasoning. SpaceXAI also introduced Grok 4 Heavy, which used parallel test-time compute to explore multiple hypotheses.

Grok 4 Fast and Grok 4.1

Grok 4 Fast focused on cost-efficient reasoning and offered a two-million-token context window in its reasoning and non-reasoning API variants. Grok 4.1, released in November 2025, concentrated on conversational quality and lower hallucination rates in information-seeking prompts.

These are separate model lines. A larger context window does not automatically mean a model is more capable, and a lower hallucination rate in one evaluation does not guarantee factual accuracy in every answer.

Grok 4.5: coding, agents, and knowledge work

Cursor announced Grok 4.5 on July 8, 2026; SpaceXAI published its full launch report on July 16. SpaceXAI describes Grok 4.5 as its frontier model for coding, agentic tasks, and knowledge work. Cursor confirms that it jointly trained the mixture-of-experts model with SpaceXAI using trillions of tokens of codebase and developer-agent interaction data, alongside broader STEM and research material.

Grok 4.5 is now the default model in Grok Build and is available through the SpaceXAI console and Cursor.

What can Grok do today?

Search current information

The API exposes Web Search and X Search as server-side tools. This is useful for news, market research, and source discovery, but search access is not the same as guaranteed truth. A reliable workflow still checks publication dates, distinguishes primary from secondary sources, and preserves citations.

Reason and use tools

Grok models can call functions, execute code, search collections, and use remote MCP tools when the application enables them. The model supplies decisions; the surrounding agent supplies permissions, tools, memory, and validation.

Work with text, images, files, and office documents

Grok 4.5 accepts text and image input and returns text. Separate Imagine and Voice models cover image, video, and speech workflows. SpaceXAI also offers Grok add-ins for Word, PowerPoint, Excel, and Outlook. These products should not be confused with the input/output modalities of the grok-4.5 API model itself.

Run as a coding agent

Grok Build is SpaceXAI's terminal coding agent. It can inspect repositories, propose a plan, edit files, run commands, and delegate work to subagents. On July 15, 2026, SpaceXAI open-sourced the Grok Build harness, including its agent loop, tools, terminal UI, and extension system. The open-source harness does not make Grok 4.5's weights open source.

Grok 4.5 API specifications

According to the current SpaceXAI developer documentation:

Knowledge cutoff and live search are different concepts. The cutoff describes information learned during training; enabled search tools can retrieve newer sources at request time.

Minimal API example

curl https://api.x.ai/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -d '{
    "model": "grok-4.5",
    "input": "Compare these two implementation plans and list the risks."
  }'

In production, keep API keys server-side, restrict tool permissions, log token and tool costs, and evaluate outputs against a task-specific test set.

Is Grok accurate?

No frontier model is a source of truth by itself. SpaceXAI reported lower hallucination rates for Grok 4.1 than for its previous non-reasoning model, but the company also advises users to verify outputs. Independent evaluator Artificial Analysis currently reports that Grok 4.5 scores 54 on its Intelligence Index and has a February 2026 knowledge cutoff.

For factual work:

1. enable search only when freshness is needed;
2. ask for primary sources and open them;
3. verify numerical claims and dates independently;
4. separate retrieved evidence from the model's inference;
5. require human approval for medical, legal, financial, security, or production-changing decisions.

Which Grok product should you choose?

Choose with an evaluation set, not a leaderboard alone. Measure task success, latency, token use, retry rate, and review time.

Frequently asked questions

Is Grok the same as xAI?

No. Grok is the model and product family. xAI was the company that created Grok; after SpaceX acquired xAI in 2026, official materials use the SpaceXAI name.

Is Grok connected to X in real time?

Grok can use X Search when the product or API request enables that tool. That gives it access to recent posts, but the result still needs source and credibility checks.

What is the latest Grok model?

As of July 23, 2026, SpaceXAI lists Grok 4.5 as its latest flagship model. Model availability can vary by product and account.

How much does Grok 4.5 cost through the API?

The standard listed price is $2 per million input tokens and $6 per million output tokens. Tools, priority service, third-party gateways, or enterprise agreements may add different charges.

Is Grok 4.5 open source?

No. Its weights and parameter count are not public. Grok Build's agent harness is open source, which is a separate matter.

Has Grok achieved AGI?

There is no accepted evidence that Grok or any other current model has achieved broadly agreed artificial general intelligence. SpaceXAI markets Grok as a frontier model; that is not an AGI certification.

Sources and update policy

• SpaceXAI: Grok 3 Beta
• SpaceXAI: Grok 4
• SpaceXAI: Grok 4 Fast
• SpaceXAI: Grok 4.1
• SpaceXAI: Introducing Grok 4.5
• SpaceXAI Developers: Grok 4.5
• Cursor: Introducing Grok 4.5
• Artificial Analysis: Grok 4.5