Grok models are the AI models that SpaceXAI (formerly xAI) develops for the Grok chatbot, the API and coding environments such as Cursor. Grok 3, announced on 19 February 2025, was followed by Grok 4, Grok 4.1, Grok 4.20, Grok 4.3, Grok 4.5, Grok 4.6 and Grok 4.7. Grok 4.7 appeared on 21 September 2026 and is the latest model.
Grok 3 and Grok 4 make up the 2025 generation. Grok 3 brought Think mode and DeepSearch, while Grok 4 added built-in tool use and gained the Grok 4 Heavy variant. In Auto mode, the Grok AI chatbot selects a model itself, and you can set a fixed version in the model picker.
After Grok 4 came a series of versions with a decimal point: Grok 4.1 in November 2025, Grok 4.20 in March 2026 and then Grok 4.3, 4.5, 4.6 and 4.7. According to SpaceXAI’s announcements, the newer models focus on coding, agents and knowledge work, while the largest context window belongs to an older model.
No Grok 5 has been announced. Eight older models, including Grok 3 and Grok 4, were removed from the API on 15 May 2026 and now redirect to Grok 4.3 or another newer model.
Which Grok models are there?
The Grok models since Grok 3 comprise ten text models and main variants, plus separate models for images, video and voice. The table sets out the date and main feature of each text model according to its announcement; smaller variants such as Grok 3 mini and Grok 4.1 Fast are covered in the section on the model itself.
| Model | Announced or live | Main feature according to the announcement |
|---|---|---|
| Grok 3 | 19 February 2025 | Reasoning model with Think mode and DeepSearch |
| Grok 4 | 9 July 2025 | Built-in tool use and real-time search |
| Grok 4 Heavy | 9 July 2025 | Works through several hypotheses in parallel |
| Grok 4 Fast | 19 September 2025 | Cost-efficient reasoning model, 2 million tokens of context |
| Grok 4.1 | 17 November 2025 | Focused on conversation, with and without a reasoning step |
| Grok 4.20 | 10 March 2026 | Reasoning model with 1 million tokens of context |
| Grok 4.3 | No date published | 1 million tokens of context, handles requests to old model names |
| Grok 4.5 | 16 July 2026 | Trained together with Cursor, focused on coding |
| Grok 4.6 | 12 August 2026 | Long-running agents, 500,000 tokens of context |
| Grok 4.7 | 21 September 2026 | Coding and knowledge work, 500,000 tokens of context |
Besides these text models, SpaceXAI has the Grok Imagine image and video models and the Grok Voice speech models, with Grok Voice 3.0 since 14 October 2026. A Fast variant is a faster or more economical version of a model, and in the API old model names live on as redirects to a newer model.
What is Grok 3?
Grok 3 is the reasoning model xAI announced on 19 February 2025, together with the more economical variant Grok 3 mini. Grok 3 introduced two modes. Think mode shows the reasoning steps towards an answer, and DeepSearch, which xAI called the company’s first agent, combines information from several sources into a report.
According to the announcement, Grok 3 was trained on Colossus, xAI’s supercomputer. X Premium and Premium+ subscribers got Grok 3 first on X and grok.com, other users followed later with usage limits, and the API arrived on 3 April 2025.
Grok 3 has not been in the API since 15 May 2026: a request to the name grok-3 goes to Grok 4.3.
What is Grok 4?
Grok 4 is the model xAI released on 9 July 2025, with built-in tool use such as a web browser and a code interpreter, and with real-time search. In the API, Grok 4 had a context window of 256,000 tokens: the amount of text, measured in tokens, that a model processes per request.
Since 15 May 2026, the API name grok-4-0709 redirects to Grok 4.3.
What is Grok 4 Heavy?
Grok 4 Heavy is the more extensive variant of Grok 4, released on 9 July 2025 alongside Grok 4. According to xAI, Grok 4 Heavy works through several hypotheses at once by applying extra computing power in parallel to each question.
What is Grok 4 Fast?
Grok 4 Fast is the variant of Grok 4 that xAI announced on 19 September 2025 as a cost-efficient reasoning model. Grok 4 Fast combined modes with and without a reasoning step in a single model, with a context window of 2 million tokens.
Grok 4.1 Fast followed on 19 November 2025, together with the Agent Tools API for developers. Both Fast models were removed from the API on 15 May 2026, in both the reasoning and non-reasoning versions.
What is Grok 4.1?
Grok 4.1 is the update of 17 November 2025 that xAI aimed at creative and emotional conversations, with, it said, fewer factual errors on information queries. Grok 4.1 came in two variants: Grok 4.1 Thinking, with a reasoning step before the answer, and a variant without a reasoning step for direct answers.
According to the announcement, Grok 4.1 became available to all users on grok.com, X and the iOS and Android apps. Grok 4.1 was rolled out in Auto mode and was available as a fixed choice in the model picker.
What is Grok 4.20?
Grok 4.20 is the reasoning model that went live on 10 March 2026, together with the Grok 4.20 Multi-agent variant. According to SpaceXAI’s API pricing table, Grok 4.20 has a context window of 1 million tokens.
What is Grok 4.3?
Grok 4.3 is the model with a 1 million token context window that has handled requests to old Grok model names in the API since 15 May 2026. A call to grok-3, grok-4-0709 or a Fast model now returns an answer from Grok 4.3, with a low reasoning level or without a reasoning step. SpaceXAI gives no release date for Grok 4.3.
What is Grok 4.5?
Grok 4.5 is the model SpaceXAI trained together with Cursor and announced on 16 July 2026. According to the announcement, Grok 4.5 focuses on coding, agents and knowledge work, and it became the default model in the Grok Build coding tool.
Grok 4.5 was rolled out in three steps. The release notes give 8 July 2026 for the API and 17 July 2026 for the API console for European users. On 22 July 2026, grok.com, X and the iOS and Android apps followed.
What is Grok 4.6?
Grok 4.6 is the successor to Grok 4.5, released on 12 August 2026 and, according to SpaceXAI, built for agents that work on a single task for a long time. Grok 4.6 has a context window of 500,000 tokens.
Besides the API and Cursor, Grok 4.6 became available in August 2026 through GitHub Copilot, Amazon Bedrock, the Gemini Enterprise Agent Platform and Microsoft Foundry.
What is Grok 4.7?
Grok 4.7 is the latest Grok model, released on 21 September 2026 and, according to SpaceXAI’s announcement, focused on coding and knowledge work. According to the same announcement, Grok 4.7 runs on a new, larger base model than Grok 4.6, with a longer training run on harder tasks.
Grok 4.7 accepts text and images as input, produces text as output and has a context window of 500,000 tokens. Developers choose from four reasoning levels: low, medium, high and xhigh.
Grok 4.7 is available in Cursor, Grok Build, the API and on cloud platforms. According to SpaceXAI, the Grok 4.7 Fast variant delivers twice the output speed and runs only in Cursor and Grok Build.
Is there going to be a Grok 5?
As of 16 October 2026, no Grok 5 had been announced: SpaceXAI’s news page and release notes mention neither a Grok 5 nor a Grok 4.8. SpaceXAI has not published a release date or specifications for Grok 5.
Which Grok models are no longer available?
According to SpaceXAI’s migration page, eight Grok models were removed from the API on 15 May 2026. Since then, every request to an old name automatically goes to a newer model:
- grok-3 and the two non-reasoning Fast models go to Grok 4.3 without a reasoning step.
- grok-4-0709 and the two reasoning Fast models go to Grok 4.3 with a low reasoning level.
- grok-code-fast-1, the coding model from August 2025, goes to Grok Build 0.1.
- grok-imagine-image-pro went to grok-imagine-image-quality; since 2 November 2026, that model has also been retired, and requests to both names go to grok-imagine-image-2.0 at low quality.
Before Grok 3, xAI released three model lines: Grok-1, Grok-1.5 and Grok-2. Grok was announced on 3 November 2023 and Grok-1 was released as an open model on 17 March 2024. Grok-1.5 followed on 28 March 2024, the vision version Grok-1.5 Vision on 12 April 2024 and Grok-2 on 13 August 2024.
What is the latest Grok model?
Grok 4.7 is the latest Grok model, available since 21 September 2026 in Cursor, Grok Build, the API and on cloud platforms. On the API models page, Grok 4.7 is the only text model. SpaceXAI did not mention the Grok app as a channel at launch.
Which Grok model should you choose for which task?
The choice of Grok model depends on the task and the amount of text, not on the highest version number. Based on what SpaceXAI documents for each model, the choice breaks down as follows:
- Coding and knowledge work: Grok 4.7, the focus of the September 2026 announcement.
- Long documents: Grok 4.3 or Grok 4.20, each with a context window of 1 million tokens.
- Long-running agents: Grok 4.6, which SpaceXAI built for tasks that run for a long time.
- Questions in the app: the model from Auto mode, or a fixed choice in the model picker.
Which Grok model has the largest context window?
Grok 4.3 and Grok 4.20 have the largest context window of the Grok models in SpaceXAI’s current API pricing table, with 1 million tokens each, although they are no longer on the models page. Grok 4.6 and Grok 4.7 process 500,000 tokens per request, half that amount. So a newer Grok model does not automatically have more room for long documents.
Grok 4 Fast had a context window of 2 million tokens at its launch in September 2025, but has not been in the API since 15 May 2026.
Which Grok model runs in the Grok app?
The Grok app selects the model automatically in Auto mode, and you can set a fixed Grok model yourself in the model picker, as xAI described at the launch of Grok 4.1. The Grok product page does not give a version number for each subscription.
If you want to use Grok for work with one fixed model version, choose that model in the model picker instead of Auto mode.
Are all Grok models available in Europe?
No, not every Grok model was available in Europe straight away at launch. Grok 4.5 was in the API on 8 July 2026, but European users only got access in the API console on 17 July 2026.
SpaceXAI does not publish a separate European date for Grok 4.7. The model picker in the app or the API console shows which Grok models are active for your account.
Frequently asked questions about Grok models
If you work with Grok models through the API or on your own hardware, you run into practical questions that go beyond the version history.
Which Grok models can you download yourself?
Two Grok models can be downloaded: Grok-1 and Grok 2, which are on the xai-org Hugging Face page. Grok-1 has 314 billion parameters and falls under the Apache 2.0 licence. Grok 2 is covered by the Grok 2 Community License Agreement, and according to the model page you need eight GPUs with more than 40 GB of memory each. That page has no weights for Grok 3 or newer models.
Which model name do you use in the API for a fixed Grok version?
For a fixed Grok version, you use the model name with a date in the API, in the form modelname-date, because it points directly to one specific release. According to SpaceXAI’s documentation, the name without a suffix points to the latest stable version, and the name with -latest to the newest version. Those two change with each update, which is handy for new features but less so for workflows that must give the same output every time.
Where does the Grok API process your requests?
The Grok API processes your requests through a global endpoint, and according to the documentation, SpaceXAI may move requests between regions for capacity or reliability. If you want to be sure that processing and inference take place in the United States, you use the US endpoint us.api.x.ai. The documentation does not mention a European region, and SpaceXAI itself states that a regional endpoint is not, on its own, a full guarantee of data residency.
What does a new Grok model mean for the way you work?
A new Grok model answers the same prompt differently from its predecessor, because a different model with different training processes the question. If you worked with Grok 3 or Grok 4 in 2025 and now use Grok 4.7, you are working with a different context window and different reasoning levels.
After every model change, check which Grok model is active and whether the output still matches the source. Client data and personal data do not belong in an AI tool without a legal basis and agreements with the processor, even when a newer model can handle more text.
The AI Literacy Course trains you to check AI output and recognise which data does not belong in an AI tool, for whichever model is active at the time.
