On July 9, 2025, Elon Musk's AI company xAI released Grok 4 and Grok 4 Heavy, a multi-agent version reserved for its most expensive subscription. xAI claimed state-of-the-art results on some of the hardest reasoning benchmarks available, and it offered Grok 4 through an API with a 256,000-token context window at mid-tier prices. The launch put xAI in the same conversation as OpenAI, Google and Anthropic for the first time, but it also arrived in the middle of a controversy over the Grok chatbot's behavior on X, which made the model a test case for how developers should weigh capability against trust.

Key Facts#

  • Release date: July 9, 2025. The API model ID is grok-4-0709, also reachable through the grok-4 alias.
  • Two variants: Grok 4, and Grok 4 Heavy, which runs several agents on the same problem and uses their combined work to answer.
  • Vendor-reported results: 15.9% on the ARC-AGI-2 reasoning benchmark, then a new high for a commercial model, and 25.4% on Humanity's Last Exam without tools. xAI reported 44.4% on Humanity's Last Exam for Grok 4 Heavy with tools.
  • API: a 256,000-token context window, priced at $3 per million input tokens and $15 per million output tokens.
  • Consumer access: the SuperGrok plan at $30 per month, plus a SuperGrok Heavy tier at $300 per month for access to Grok 4 Heavy.
  • Early integrations: coding and agent tools including Cursor, Cline and LangChain, as well as Perplexity's paid plans, added Grok 4 around launch.
  • Enterprise channel: Microsoft's Foundry model catalog now includes grok-4, with access that requires registration.

What Happened#

xAI announced the models on its website, and the headline claim was reasoning performance. On ARC-AGI-2, a reasoning benchmark designed to resist memorization, Grok 4's 15.9% was reported at the time as roughly double the best previous score from a commercial model. On Humanity's Last Exam, a collection of expert-level questions across many fields, xAI reported 25.4% for Grok 4 without tools and 44.4% for Grok 4 Heavy with tools. xAI also highlighted a leading result on Vending-Bench, a simulation in which an agent runs a small vending business over many steps, a test of long-horizon consistency rather than single answers.

Grok 4 Heavy is the more unusual product. It spends more compute at inference time, running multiple agents in parallel on the same task and combining their results. That approach, sometimes called parallel test-time compute, trades cost for accuracy, and xAI tied it to a subscription tier priced at ten times its standard plan.

For developers, the API mattered more than the benchmarks. Grok 4 launched with a 256,000-token context window and pricing of $3 and $15 per million input and output tokens, placing it in the same bracket as mid-tier frontier models rather than the most expensive flagships. Tool vendors moved quickly: AI newsletter coverage at launch noted integrations in Cursor, Cline, LangChain and Perplexity.

The launch did not happen in a vacuum. As the AI testing company promptfoo later put it, Grok 4 launched "amid Hitler-praising controversies" involving the Grok chatbot on X. The episode shaped coverage of the release and raised questions about how xAI tests and governs model behavior.

Background#

Newsletter coverage of the launch framed Grok 4 as xAI going from nothing to a state-of-the-art model in about two years. Grok 4 succeeded the Grok 3 family as xAI's flagship model.

The timing was competitive. OpenAI had shipped o3 and o4-mini in April, Anthropic had released Claude Opus 4 and Sonnet 4 in May, and Google had made Gemini 2.5 Pro generally available in June. Grok 4 was xAI's bid to be counted among the frontier labs rather than as a chatbot attached to a social network.

Why It Matters for Developers#

xAI's API follows the OpenAI request format, so .NET teams can trial Grok with the official OpenAI library by pointing it at xAI's endpoint, and then use it behind the same IChatClient abstraction as every other provider:

C#
using System.ClientModel;
using Microsoft.Extensions.AI;
using OpenAI;
using OpenAI.Chat;

string apiKey = Environment.GetEnvironmentVariable("XAI_API_KEY")
    ?? throw new InvalidOperationException("Set XAI_API_KEY.");

var grok = new ChatClient(
    model: "grok-4-0709",
    credential: new ApiKeyCredential(apiKey),
    options: new OpenAIClientOptions { Endpoint = new Uri("https://api.x.ai/v1") });

IChatClient client = grok.AsIChatClient();
var response = await client.GetResponseAsync("Review this LINQ query for N+1 problems: ...");
Console.WriteLine(response.Text);

Compatibility at the wire level is not the same as identical behavior, though. Tool-calling quirks, refusal patterns and output formatting differ between providers, so run your own evaluation suite before routing production traffic. Our guide to evaluating AI applications in .NET covers building one.

Grok 4 Heavy is also a useful design lesson. You do not need a $300 subscription to use parallel test-time compute: an application can fan out the same prompt to several model calls, then use a judge step or majority vote to pick an answer. That pattern, described in our guide to AI agent architecture patterns, is worth its cost only for high-value decisions, so measure accuracy gains against the extra tokens.

Finally, the controversy around the launch is a reminder that model choice is also a governance decision. Promptfoo published an open methodology and a dataset of 2,500 political questions to measure bias, and its July 2025 analysis found Grok more right-leaning than most other models it tested but still left of center. Whatever model you pick, put output filtering, logging and human review around it, as described in our guide to responsible AI and LLM security.

What's Next#

xAI did not stand still after Grok 4. Its later releases included faster Grok 4.1 variants and a Grok 4.20 family with reasoning, non-reasoning and multi-agent versions, and by September 2026 Microsoft's Foundry catalog listed models up to Grok 4.6 in preview alongside the original grok-4. Microsoft now lists these models under the SpaceXAI name, reflecting xAI's combination with SpaceX, which we cover in our SpaceX and xAI merger story. For .NET teams on Azure, that catalog is the most straightforward enterprise route to the Grok models, with Azure billing and governance.

The open questions from the Grok 4 launch remain relevant. Benchmarks such as ARC-AGI-2 and Humanity's Last Exam moved quickly after July 2025, so a record score is a snapshot, not a durable advantage. How consistently xAI governs model behavior is the other variable. Teams evaluating Grok models should weigh both, and keep the provider swappable.

Sources#