xAI Grok 4 Release Notes

xAI has announced the release of Grok 4, a model designed to push the boundaries of frontier intelligence through scaled reinforcement learning and native tool use. The model is available to SuperGrok and Premium+ subscribers, as well as via the xAI API.

Scaled Reinforcement Learning and Infrastructure

Grok 4's reasoning capabilities were developed by scaling reinforcement learning (RL) training to pretraining scale. This advancement was supported by the Colossus 200,000 GPU cluster and infrastructure and algorithmic innovations that increased compute efficiency by 6x.

To improve the model's versatility, xAI expanded verifiable training data beyond math and coding into multiple additional domains. This approach resulted in smooth performance gains while utilizing over an order of magnitude more compute than previous training runs.

Native Tool Use and X Integration

Grok 4 is trained with RL to natively use tools, allowing it to augment its reasoning with a code interpreter and web browsing. The model autonomously determines its own search queries to synthesize high-quality responses from across the web.

Integration with the X platform is a core feature, enabling Grok 4 to use advanced keyword and semantic search tools, as well as media analysis, to retrieve real-time information from within X.

Grok 4 Heavy and Parallel Test-Time Compute

Grok 4 Heavy is a specialized version of the model that utilizes parallel test-time compute, allowing it to consider multiple hypotheses simultaneously. This variant sets new performance standards for reliability and complex reasoning:

Humanity's Last Exam: Grok 4 Heavy is the first model to score 50% (specifically 50.7% on the text-only subset) on this expert-level benchmark.
Academic Benchmarks: The model saturates most academic benchmarks.

Frontier Intelligence Benchmarks

Grok 4 demonstrates state-of-the-art performance across several high-difficulty benchmarks for closed models:

ARC-AGI V2: Achieves 15.9%, nearly double the ~8.6% of Claude Opus.
Vending-Bench (Agentic): Reaches a net worth of $4,694.15 and 4,569 units sold, significantly outperforming Claude Opus 4 ($2,077.41) and humans ($844.05).
USAMO'25: Grok 4 Heavy leads with a score of 61.9%.

API Capabilities and Enterprise Security

The Grok 4 API provides developers with multimodal understanding and a 256,000 token context window. Key features include:

Live Search API: Integrates real-time data search across X, the web, and news sources.
Compliance: The API is SOC 2 Type 2, GDPR, and CCPA certified.
Availability: The model is coming soon to hyperscaler partners for enterprise deployment.

Multimodal Voice and Vision Mode

xAI has introduced an upgraded Voice Mode featuring enhanced realism and responsiveness. This mode includes a new voice and a redesigned conversation interface.

Additionally, Grok 4 can now process live visual input. By enabling video during a voice chat, the model can analyze the user's camera feed in real-time and provide insights based on what it sees. This capability was developed using an in-house RL framework and speech compression techniques.

Future Roadmap

xAI intends to continue scaling reinforcement learning to move beyond verifiable rewards in controlled domains toward solving complex, real-world problems in dynamic environments. Future development will focus on improving multimodal integration of vision and audio to create more intuitive and efficient AI systems.