Grok 4 is a large language model developed by xAI and released on July 9, 2025. Designed for reasoning and tool-assisted problem solving, it introduced integrated search and code-execution capabilities. The launch also included Grok 4 Heavy, a variant using additional computation to consider multiple hypotheses in parallel. Initially, access was offered through paid consumer subscriptions and an application programming interface (API). (x.ai)
Development and training
Grok 4 extended xAI’s work on reasoning models through substantially scaled-up reinforcement learning. According to the company, training used its Colossus cluster of approximately 200,000 graphics processing units, with verifiable tasks expanded beyond mathematics and programming. (x.ai)
The August 2025 model card describes training data drawn from publicly available Internet material, third-party providers, users or contractors, and internally generated content. Data preparation included deduplication and classification. Post-training combined human feedback, verifiable rewards, model-based grading, and supervised fine-tuning for particular capabilities. (data.x.ai)
Reasoning and tool use
Grok 4 was trained to incorporate tools into its problem-solving process, including a code interpreter, web browsing, and keyword and semantic search across X. It could formulate search queries and retrieve external information while developing an answer, rather than relying exclusively on information encoded during training. (x.ai)
This tool-calling capability also supported AI-agent applications: systems in which the model repeatedly selects actions toward a goal. The model card explicitly distinguishes ordinary conversation from agentic operation involving real function calls, because the latter introduces additional opportunities for misuse or task hijacking. (data.x.ai)
Variants and subsequent development
Grok 4 Heavy
Grok 4 Heavy applied additional parallel test-time computation, allowing multiple hypotheses to be considered simultaneously. It was introduced with the SuperGrok Heavy subscription tier. This distinguished it from the standard release by the computation applied while answering, not simply by subscription branding. (x.ai)
Grok 4 Fast
Released on September 19, 2025, Grok 4 Fast was a subsequent efficiency-focused model informed by the development of Grok 4. It offered a two-million-token context window and a unified architecture supporting reasoning and non-reasoning modes. The API exposed these modes through separate model identifiers, grok-4-fast-reasoning and grok-4-fast-non-reasoning. (x.ai)
Its model card describes lower latency and cost than the original Grok 4, with the option to omit extended reasoning for latency-sensitive applications. These characteristics made Fast a distinct deployment option rather than another name for the original July release. (data.x.ai)
Safety evaluation and limitations
The Grok 4 model card, updated on August 20, 2025, evaluated three broad areas: abuse potential, concerning behavioral tendencies, and dual-use capabilities. Reported safeguards included refusal instructions, warnings against adversarial prompts, and model-based input filters. Tests covered harmful requests, agentic misuse, deceptive behavior, and task hijacking through injected instructions. (data.x.ai)
The reported results were not uniformly zero-risk. For example, xAI reported a 2% attack-success rate on its AgentDojo task-hijacking evaluation and a 43% dishonesty rate on the MASK evaluation under the tested conditions. These are results for particular evaluation procedures, not measurements of the frequency of errors or dishonesty across all real-world conversations. (data.x.ai)
API retirement and version identification
The original dated API release was identified as grok-4-0709. xAI retired that model on May 15, 2026, at 12:00 p.m. Pacific Time, alongside several other earlier models, including the Grok 4 Fast reasoning and non-reasoning endpoints. Its migration notice states that requests to those retired identifiers automatically redirect to grok-4.3; reasoning-model requests use the replacement model’s low reasoning-effort setting. (docs.x.ai)
Consequently, continued acceptance of the original API identifier does not establish that the July 2025 model is still serving a request. The distinction matters for reproducibility, because the retained identifier can invoke a different model with different behavior and pricing. (docs.x.ai)
References
- Grok 4x.ai
- Grok 4 Model Carddata.x.ai
- Grok 4 Fastx.ai
- Grok 4 Fast Model Carddata.x.ai
- Grok Model Retirement on May 15, 2026docs.x.ai